U.S. patent application number 10/966508 was filed with the patent office on 2005-03-31 for accelerating a distributed component architecture over a network using a direct marshaling.
This patent application is currently assigned to Microsoft Corporation. Invention is credited to Forin, Alessandro, Hunt, Galen C., Wang, Yi-Min.
Application Number | 20050071857 10/966508 |
Document ID | / |
Family ID | 22340460 |
Filed Date | 2005-03-31 |
United States Patent
Application |
20050071857 |
Kind Code |
A1 |
Wang, Yi-Min ; et
al. |
March 31, 2005 |
Accelerating a distributed component architecture over a network
using a direct marshaling
Abstract
A method for improving the performance of a distributed object
model over a network is disclosed. A client computer contains a
client object which can call an interface on a server object
located on a server computer. Rather than copying all of the call
parameters into an RPC buffer for transmission across the network,
a network interface card with scatter-gather capability can be
used. The RPC data can contain only a list of pointers into the
client memory and a size of each parameter. The network interface
card can then grab the parameters directly from the client memory
using the list in the RPC buffer without the need to copy the data
itself. At the server side, the network interface card can place
the parameters into an RPC buffer, or if the size is known
beforehand, directly into the server memory. The server can also
access the parameters directly from the RPC buffer. On the return,
the server can use a callback function to indicate when its network
interface card has finished sending the response data so that the
server does not clear its memory prematurely. At the client side,
if the size of the response is not known, and the data is placed
into the RPC buffers, it can be copied from the RPC buffer into the
client memory.
Inventors: |
Wang, Yi-Min; (Bellevue,
WA) ; Hunt, Galen C.; (Bellevue, WA) ; Forin,
Alessandro; (Bellevue, WA) |
Correspondence
Address: |
CHRISTENSEN, O'CONNOR, JOHNSON, KINDNESS, PLLC
1420 FIFTH AVENUE
SUITE 2800
SEATTLE
WA
98101-2347
US
|
Assignee: |
Microsoft Corporation
Redmond
WA
|
Family ID: |
22340460 |
Appl. No.: |
10/966508 |
Filed: |
October 14, 2004 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
10966508 |
Oct 14, 2004 |
|
|
|
09458139 |
Dec 9, 1999 |
|
|
|
6826763 |
|
|
|
|
60111788 |
Dec 11, 1998 |
|
|
|
Current U.S.
Class: |
719/330 |
Current CPC
Class: |
G06F 9/465 20130101;
G06F 9/547 20130101 |
Class at
Publication: |
719/330 |
International
Class: |
G06F 009/46 |
Claims
We claim:
1. A method of communication between a first object located on a
first computer and a second object located on a second computer,
the second computer having a memory storage location and a buffer,
the first and second computers connected by a network, accessed by
the second computer through a network interface card on the second
computer, the method comprising: receiving a call from the first
object on an interface of the second object; receiving, by the
network interface card, a parameter of the call from the first
object; storing the parameter in a memory location; and accessing,
by the second object, the parameter.
2. The method of claim 1 wherein the memory location is the
buffer.
3. The method of claim 2 wherein the accessing the parameter is
performed in the buffer.
4. The method of claim 2 further comprising copying the parameter
from the buffer into the memory storage location, wherein the
accessing the parameter is performed in the memory storage
location.
5. The method of claim 1 wherein the memory location is the memory
storage location, and wherein the accessing the parameter is
performed in the memory storage location.
6. The method of claim 1 wherein the receiving comprises: storing,
on the second computer, a second data into a first receive buffer,
wherein the first receive buffer is posted prior to sending a first
data to the first computer.
7. The method of claim 6 wherein the first data to the first
computer is sent prior to receiving the second data from the first
computer.
8. The method of claim 6 wherein the receiving further comprises:
cleaning up, on the second computer, a send buffer after sending
the first data to the first computer and prior to receiving the
second data from the first computer.
9. The method of claim 8 wherein the send buffer is used to send
the first data to the first computer.
10. The method of claim 6 wherein the receiving further comprises:
cleaning up, on the second computer, a second receive buffer after
sending the first data to the first computer and prior to receiving
the second data from the first computer.
11. A computer-readable medium having computer-executable
instructions for performing steps for communicating between a first
object located on a first computer and a second object located on a
second computer, the second computer having a memory storage
location and a buffer, the first and second computers connected by
a network, accessed by the second computer through a network
interface card on the second computer, the steps comprising:
receiving a call from the first object on an interface of the
second object; receiving, by the network interface card, a
parameter of the call from the first object; storing the parameter
in a memory location; and accessing, by the second object, the
parameter.
12. The computer-readable medium of claim 11 wherein the memory
location is the buffer.
13. The computer-readable medium of claim 12 wherein the accessing
the parameter is performed in the buffer.
14. The computer-readable medium of claim 12 having further
computer-executable instructions for performing steps comprising:
copying the parameter from the buffer into the memory storage
location, wherein the accessing the parameter is performed in the
memory storage location.
15. The computer-readable medium of claim 11 wherein the memory
location is the memory storage location, and wherein the accessing
the parameter is performed in the memory storage location.
16. The computer-readable medium of claim 11 wherein the receiving
comprises: storing, on the second computer, a second data into a
first receive buffer, wherein the first receive buffer was posted
prior to sending a first data to the first computer.
17. The computer-readable medium of claim 16 wherein the first data
to the first computer is sent prior to receiving the second data
from the first computer.
18. The computer-readable medium of claim 16 wherein the receiving
further comprises: cleaning up, on the second computer, a send
buffer after sending the first data to the first computer and prior
to receiving the second data from the first computer.
19. The computer-readable medium of claim 18 wherein the send
buffer is used to send the first data to the first computer.
20. The computer-readable medium of claim 20 wherein the receiving
further comprises: cleaning up, on the second computer, a second
receive buffer after sending the first data to the first computer
and prior to receiving the second data from the first computer.
Description
CROSS-REFERENCE TO RELATED APPLICATION
[0001] This application is a continuation of prior U.S. patent
application Ser. No. 09/458,139, filed Dec. 9, 1999, (allowed on
Aug. 16, 2004) which claims the benefit of U.S. Provisional
Application Ser. No. 60/111,788 filed Dec. 11, 1998, both of which
are incorporated herein by reference.
TECHNICAL FIELD
[0002] This invention relates generally to software communication
over a network and, more particularly, relates to acceleration of
the interaction of objects over a network.
BACKGROUND OF THE INVENTION
[0003] A component object model defines the interactions between
computer software components. The advantage of component
programming is that it facilitates the use of reusable sections of
code. Programs will often provide similar functionality. For
example, many modern software applications provide pull-down menu
functionality. Computer code that allows a user to pull down a menu
on the computer screen can be found in some form in each of these
applications. A component providing the same functionality,
however, would only need to be written once, and then simply reused
by each succeeding application. The time required to create an
application, therefore, can be significantly reduced by reusing
preexisting components.
[0004] For object-based component programming to be successful, a
standard method of interactions between objects must be defined.
One such standard is the Component Object Model, or COM. COM
mandates that all objects interact through interfaces. Each
interface is a collection of functions that the object can perform.
The object is said to have "exposed" the methods contained in its
interfaces, which can then be "called", or used, by another object.
Another standard, based on COM is the Distributed Component Object
Model, or DCOM. DCOM defines a standard method of interaction
between objects that may be located on remote computers connected
through a network. DCOM uses a Remote Procedure Call (RPC) model to
define a method of communication between objects across a network.
The RPC model is independent of the underlying network structure or
protocols.
[0005] As can be expected, calling an object located on the same
computer is faster than calling an object located on a remote
computer. This speed difference can be due to a number of factors.
The network cables are significantly longer than the leads between
the processor and the memory on the local machine. Therefore, the
electrical signals simply take longer to reach the remote computer
than to reach the object resident in memory on the local machine. A
significantly larger factor is the overhead caused by the network
protocol. Each data transmission over a network must be
encapsulated, and additional information must be added to the
packet so that it may be transferred across the network with error
correcting capabilities, and so that it may properly be decoded on
the remote machine. Furthermore, each packet sent over a network
may be accompanied by a flurry of additional network packets
performing necessary buffer management and receipt acknowledge
functions. These further packets, which comprise the network flow
control, also add to the time required to send an object call over
a network to a remote computer.
[0006] An additional factor contributing to the speed difference
between a call to an object resident on the same machine and one
resident on a remote machine is the overhead created by DCOM and
the RPC model. RPC marshals pointers and data to be transmitted
across the network by reading them from the program memory and
packaging them for transportation across the network. Marshaling
introduces delay because it copies from program memory into an RPC
buffer the element that is to be transmitted across the network.
Another aspect of the overhead of DCOM and RPC are the runtime
layers. The RPC and DCOM runtime layers bridge together the client
and server so that the client can make remote calls to the server.
This process of bridging the client and server together is known as
binding. Binding information can include the Internet Protocol (IP)
address, the port number, and the interface identifier (IID).
[0007] The combined effects of the marshaling, the additional
packets of flow control, and the activities of the runtime layers
result in a dramatic decrease in the performance of DCOM over a
network. In fact, compared to a raw network application which
directly sends data across the network, the DCOM overhead can
decrease performance by a factor of three or more.
SUMMARY OF THE INVENTION
[0008] Accordingly, the present invention provides a method for
increasing the efficiency of calling remote objects over a network
using DCOM.
[0009] The present invention also provides a more efficient method
of marshaling DCOM application data.
[0010] The invention additionally provides a method for maximizing
the efficiency of RPC flow control.
[0011] The invention also provides for a more efficient binding
between the client and the server.
[0012] An object model, such as DCOM, can allow communication
across a network by making the network communication transparent to
the client and server objects. DCOM uses a "proxy" on the client
process and a "stub" on the server process to achieve such
transparency. The proxy acts as a local version of the server
object which the client can call, and the stub acts as a local
client object on the server. The proxy and stub then communicate
with one another across the network. To perform this communication,
the proxy marshals the call parameters into an RPC buffer, from
which they are transferred across the network to the stub. The stub
unmarshals the call parameters, and calls the server object
directly. Similarly, on the return, the stub marshals the call
results into an RPC buffer for transmission across the network to
the proxy, which unmarshals the results and returns them to the
client process.
[0013] The present invention allows DCOM systems using a network
interface card (NIC) with "scatter-gather" ability to gather
elements from various memory locations to avoid copying the call
parameters into the RPC buffer. Instead, the proxy or stub simply
create a pointer list in the buffer, which is then accessed by the
NIC, which can collect the elements from memory and is responsible
for sending the data across the network. To indicate that the RPC
buffer contains only a list and not the actual values themselves,
the proxy or stub can set a flag, which is understood by the NIC or
RPC runtime. On the server side, the stub code can hold onto the
buffer and not clear it until the NIC has finished sending the
data. In such a case, a callback function can be used by the NIC to
indicate that it has completed sending the data.
[0014] Another method of improving the network performance of an
object model such as DCOM is to make more efficient the
communication between a client and a server through the RPC layer.
DCOM was designed to take advantage of the existing architecture of
RPC, provides a mechanism for making calls to remote computers
connected by a network. When a local RPC object seeks to call a
remote RPC interface, the call can specify the IP address, the port
number and the RPC IID. DCOM takes advantage of the RPC structure,
except that DCOM uses an interface pointer identifier (IPID) to
uniquely specify the COM interface to which the call is being made.
In order to use the RPC structure, the DCOM client object must send
an RPC IID to the RPC runtime layer and an IPID to the DCOM runtime
layer. Because the IPID is more specific than the RPC IID, the RPC
IID is redundant and the additional computation performed by the
RPC layer is wasted.
[0015] The present invention removes the additional computation and
communication performed by RPC and allows the DCOM client to send
only an IPID. The RPC dispatching layer on the server side is
removed from the critical path. All incoming DCOM calls are
forwarded to the DCOM dispatching layer directly. The client side
can then be modified, so that the calling DCOM object only needs to
send an IPID. The removal of the RPC dispatching allows DCOM
communication to proceed without a duplication of effort, and
therefore more efficiently.
[0016] Yet another method of improving DCOM performance involves
modifying the flow control performed by the software protocol
stacks. When transmitting data, a buffer on the receiving side must
be made available before each packet of data can be sent.
Furthermore, the sender must know that the receiver has made a
buffer available, using some form of flow control, before sending a
message. With traditional transport layers, the sender waited for
an explicit "OK TO SEND" flow-control message, thereby insuring
that the receiver had sufficient resources to accept the data the
sender was waiting to transmit. In the worst case, which can be
typical for RPC and DCOM communication, the sending of each data
packet requires the sending of one flow control packet, flow
control packets account for one half of the network traffic. More
importantly, the waiting computer does no useful work while waiting
for the "OK TO SEND" flow control message. Such idle time reduces
the efficiency of the overall system.
[0017] The present invention modifies the RPC transport layer to
use an implicit flow control. Implicit flow control does not
require an explicit communication from the receiver indicating it
is ready to receive; such as an "OK TO SEND" message. Rather,
implicit flow control insures that the receiver is ready to receive
by implicitly associating flow control messages with regular
application messages. The present invention allows a sending
computer to pre-post a receive buffer prior to sending out any data
that may cause a response message to be sent from the receiving
computer. Therefore, when the receiving computer receives the data
from the sending computer, it is an implicit acknowledgement that
the sending computer is ready to receive. By pre-posting the
receive buffer prior to sending any data, the sending of data
becomes an indication that the next receive buffer is ready. Thus,
the regular application messages can be considered flow control
messages. Such a system eliminates the overhead due to the standard
flow control by relying on the request/reply semantics of RPC
communication. Additionally, the flow control of the present
invention minimizes the idle time of the sending and receiving
computers. By removing the explicit flow control messages, the
present invention allows computers to reorganize their send and
receive cycles to minimize idle time, and thereby maximize
efficiency. Note that the present invention is more efficient than
prior systems, such as the standard TCP protocol, which piggyback
explicit flow-control messages on outgoing application messages as
often as possible. For example, the TCP heuristics to piggyback
explicit flow control-messages fail to optimize flow-control in
request-reply traffic between client and server, which is exactly
the traffic for RPC, DCOM, and HTTP.
[0018] Additional features and advantages of the invention will be
made apparent from the following detailed description of
illustrative embodiments which proceeds with reference to the
accompanying figures.
BRIEF DESCRIPTION OF THE DRAWINGS
[0019] While the appended claims set forth the features of the
present invention with particularity, the invention, together with
its objects and advantages, may be best understood from the
following detailed description taken in conjunction with the
accompanying drawings of which:
[0020] FIG. 1 is a block diagram generally illustrating an
exemplary computer system on which the present invention
resides;
[0021] FIG. 2 is a block diagram generally illustrating the
operation of DCOM over an exemplary network and computers;
[0022] FIG. 3A is a block diagram generally illustrating the layers
of DCOM on a client and the transfer of data from a client to a
server;
[0023] FIG. 3B is a block diagram generally illustrating the layers
of DCOM on a server and the transfer of data from a client to a
server;
[0024] FIG. 4A is a block diagram generally illustrating the layers
of DCOM on a client and the transfer of data from a client to a
server according to one aspect of the present invention;
[0025] FIG. 4B is a block diagram generally illustrating the layers
of DCOM on a server and the transfer of data from a client to a
server according to one aspect of the present invention;
[0026] FIG. 5A is a flow chart generally illustrating the layers of
DCOM on a client and the transfer of data from a server to a client
according to one aspect of the present invention;
[0027] FIG. 5B is a flow chart generally illustrating the layers of
DCOM on a server and the transfer of data from a server to a client
according to one aspect of the present invention;
[0028] FIG. 6 is a block diagram generally illustrating the
operation of RPC dispatching and DCOM dispatching;
[0029] FIG. 7 is a communication flow diagram generally
illustrating explicit flow control;
[0030] FIG. 8 is a communication flow diagram generally
illustrating implicit flow control according to one aspect of the
present invention; and
[0031] FIG. 9 is a communication flow diagram generally
illustrating implicit flow control according to another aspect of
the present invention.
DETAILED DESCRIPTION OF THE INVENTION
[0032] Turning to the drawings, wherein like reference numerals
refer to like elements, the invention is illustrated as being
implemented in a suitable computing environment. Although not
required, the invention will be described in the general context of
computer-executable instructions, such as program modules, being
executed by a personal computer. Generally, program modules include
routines, programs, objects, components, data structures, etc. that
perform particular tasks or implement particular abstract data
types. Moreover, those skilled in the art will appreciate that the
invention may be practiced with other computer system
configurations, including hand-held devices, multi-processor
systems, microprocessor based or programmable consumer electronics,
network PCs, minicomputers, mainframe computers, and the like. The
invention may also be practiced in distributed computing
environments where tasks are performed by remote processing devices
that are linked through a communications network. In a distributed
computing environment, program modules may be located in both local
and remote memory storage devices.
[0033] With reference to FIG. 1, an exemplary system for
implementing the invention includes a general purpose computing
device in the form of a conventional personal computer 20,
including a processing unit 21, a system memory 22, and a system
bus 23 that couples various system components including the system
memory to the processing unit 21. The system bus 23 may be any of
several types of bus structures including a memory bus or memory
controller, a peripheral bus, and a local bus using any of a
variety of bus architectures. The system memory includes read only
memory (ROM) 24 and random access memory (RAM) 25. A basic
input/output system (BIOS) 26, containing the basic routines that
help to transfer information between elements within the personal
computer 20, such as during start-up, is stored in ROM 24. The
personal computer 20 further includes a hard disk drive 27 for
reading from and writing to a hard disk 60, a magnetic disk drive
28 for reading from or writing to a removable magnetic disk 29, and
an optical disk drive 30 for reading from or writing to a removable
optical disk 31 such as a CD ROM or other optical media.
[0034] The hard disk drive 27, magnetic disk drive 28, and optical
disk drive 30 are connected to the system bus 23 by a hard disk
drive interface 32, a magnetic disk drive interface 33, and an
optical disk drive interface 34, respectively. The drives and their
associated computer-readable media provide nonvolatile storage of
computer readable instructions, data structures, program modules
and other data for the personal computer 20. Although the exemplary
environment described herein employs a hard disk 60, a removable
magnetic disk 29, and a removable optical disk 31, it will be
appreciated by those skilled in the art that other types of
computer readable media which can store data that is accessible by
a computer, such as magnetic cassettes, flash memory cards, digital
video disks, Bernoulli cartridges, random access memories, read
only memories, and the like may also be used in the exemplary
operating environment.
[0035] A number of program modules may be stored on the hard disk
60, magnetic disk 29, optical disk 31, ROM 24 or RAM 25, including
an operating system 35, one or more application programs 36, other
program modules 37, and program data 38. A user may enter commands
and information into the personal computer 20 through input devices
such as a keyboard 40 and a pointing device 42. Other input devices
(not shown) may include a microphone, joystick, game pad, satellite
dish, scanner, or the like. These and other input devices are often
connected to the processing unit 21 through a serial port interface
46 that is coupled to the system bus, but may be connected by other
interfaces, such as a parallel port, game port or a universal
serial bus (USB). A monitor 47 or other type of display device is
also connected to the system bus 23 via an interface, such as a
video adapter 48. In addition to the monitor, personal computers
typically include other peripheral output devices, not shown, such
as speakers and printers.
[0036] The personal computer 20 may operate in a networked
environment using logical connections to one or more remote
computers, such as a remote computer 49. The remote computer 49 may
be another personal computer, a server, a router, a network PC, a
peer device or other common network node, and typically includes
many or all of the elements described above relative to the
personal computer 20, although only a memory storage device 50 has
been illustrated in FIG. 1. The logical connections depicted in
FIG. 1 include a local area network (LAN) 51, a wide area network
(WAN) 52, and a high-speed system area network (SAN) 56. Such
networking environments are commonplace in offices, enterprise-wide
computer networks, intranets and the Internet.
[0037] When used in a LAN networking environment, the personal
computer 20 is connected to the local network 51 through a network
interface or adapter 53. When used in a WAN networking environment,
the person computer 20 typically includes a modem 54 or other means
for establishing communications over the WAN 52. The modem 54,
which may be internal or external, is connected to the system bus
23 via the serial port interface 46. When connected to a SAN, the
personal computer 20 is connected via a high-speed network
interface 55. In a networked environment, program modules depicted
relative to the personal computer 20, or portions thereof, may be
stored in the remote memory storage device. It will be appreciated
that the network connections shown are exemplary and other means of
establishing a communications link between the computers may be
used.
[0038] In the description that follows, the invention will be
described with reference to acts and symbolic representations of
operations that are performed by one or more computers, unless
indicated otherwise. As such, it will be understood that such acts
and operations, which are at times referred to as being
computer-executed, include the manipulation by the processing unit
of the computer of electrical signals representing data in a
structured form. This manipulation transforms the data or maintains
it at locations in the memory system of the computer, which
reconfigures or otherwise alters the operation of the computer in a
manner well understood by those skilled in the art. The data
structures where data is maintained are physical locations of the
memory that have particular properties defined by the format of the
data. However, while the invention is being described in the
foregoing context, it is not meant to be limiting as those of skill
in the art will appreciate that various of the acts and operation
described hereinafter may also be implemented in hardware.
[0039] In accordance with the invention, the interactions of a
distributed object model are shown in FIG. 2. An object model can
define a standard set of rules governing the interaction of
"objects", such as objects 74 and 76. An object, as is known by
those skilled in the art, is a computer program element comprising
computer readable instructions and computer readable data. Objects
can be very useful in the programming arts because users can use
previously programmed objects to create an application, instead of
writing all of the code themselves. Objects, therefore, allow for
efficient code reuse. Once an object is created to perform specific
tasks, any user can use that object to perform those tasks. Thus,
to implement common functionality across different computer
applications, the software author need only create an object with
that functionality once, or even find an object created by another,
and then simply use it in each application.
[0040] An object model, such as the Component Object Model (COM)
from Microsoft Corporation, seeks to define a set of standards so
that objects written by one individual can be used by all, without
any changes to the object, and even without understanding how the
object is implemented internally. To accomplish this, object models
can require objects to implement interfaces. As will be known by
one of skill in the art, an interface, such as interface 82 of
object 76, is a set of functions provided by the object which a
client of the object can request. A client of an object can only
interact with an object through that object's interfaces.
Therefore, if an object is capable of performing a certain task, a
function for that task will exist in one of the object's
interfaces. To request that the object perform a task, the client
of the object can select an interface and make a call to the
appropriate function. Because clients can only interact with an
object through its interfaces, the specific structure and internal
workings of the object are irrelevant.
[0041] Due to the increasing popularity of networked computing, it
is desirable that objects can interact with one another over a
network. To maintain compatibility, and not require rewriting, an
object model which seeks to allow implementation across a network
can provide some mechanism for transparent network communication.
As seen from the client's perspective, there should be no
difference between calling an object on a local computer and an
object on a remote, networked computer. One such object model, the
Distributed Component Object Model (DCOM) from Microsoft
Corporation, uses a "proxy" in the client process and a "stub" in
the server process to achieve such transparency.
[0042] Turning again to FIG. 2, an interface 82 is shown exposed by
server object 76. As can be seen, server object 76 resides on
server computer 72. Server computer 72 and client computer 70 are
connected through a network connection 90, which can be a SAN 56.
In order to allow client object 74, resident on client computer 70,
to call interface 82 of the server object 76, DCOM creates a proxy
78 on the client computer and a stub 80 on the server computer. The
proxy 78 on the client computer 70 acts as the client-side
representative of the server object 76 by exposing an interface 84
analogous to interface 82 exposed by the server object. The client
object 74 can therefore call interface 84 in exactly the same
manner it would call any local interface. Once the client object 74
calls interface 84, the proxy 78 transfers the call across the
network connection 90 to the stub 80. It is the stub 80 which then
calls the actual interface 82 of the server object 76. The server
object 76 can respond in an analogous manner by sending the reply
data to the stub 80 and having the stub transfer it across the
network connection 90 to the proxy 78, which then presents the data
to the client object 74. Therefore, because the proxy 78 and the
stub 80 are both located on the same machine as the objects with
which they communicate, they allow those objects to make calls and
return data in exactly the same manner as they did in a local
context. The proxy 78 and stub 80 then package the calls and data
and send them across the network, allowing for a distributed object
system. The proxy 78 and stub 80, therefore, make the network
transparent to client object 74 and server object 76.
[0043] A known layered architecture of DCOM is shown in FIGS. 3A
& 3B. As shown in FIG. 3A, when the DCOM client 120 makes a
remote call, the marshaling layer 122 prepares the call for
transmission across the network. The proxy 78 acts as the
marshaling layer 122 on the client 70 and the stub 80 acts as the
marshaling layer 140 on the server 72. Marshaling, as will be
explained in more detail below, is the packaging of the call for
transmission across the network connection 90. After the marshaling
layer 122 marshals the call, the DCOM run-time layer 124 transmits
the call over the network. The DCOM run-time layer 124 transmits
the call by invoking the functionality of the Remote Procedure Call
(RPC) run-time layer 126. DCOM's structure allows it to take
advantage of the existing structure of RPC, as will be explained
below. The loadable transport layer 128 allows the systems above it
to run on any number of different protocols, which are implemented
in the protocol stacks 130. The protocol stacks 130 then create the
packets sent out over the network connection 90. On the server
side, shown in FIG. 3B, the protocol stacks 132 receive the
packets, translate them, and pass them to the loadable transport
layer 134, which further translates the data into a form which the
RPC run-time layer 136 can accept. The RPC run-time layer 136 then
accepts the communication, and creates a binding if the call is the
first call to an interface. The DCOM run-time layer 138 accepts the
communication from the RPC run-time layer 136 and passes it to the
stub 80 at the marshaling layer 140. The stub 80 unmarshals the
parameters of the call and initiates the call to the DCOM server
object 142.
[0044] The present invention provides for a number of improvements
in this layered architecture; removing overhead, and increasing the
speed of DCOM over the network connection 90. One such improvement
can be made at the marshaling layers 122 and 140. Marshaling is the
process of arranging the data provided by the DCOM client 120 when
it calls the DCOM server object 142 into a proper form for
transmission across the network connection 90. Simple data types
such as characters and integers are not difficult to marshal.
However, DCOM often passes more complex parameters, such as arrays
of information. RPC can use a standard format called the Network
Data Representation (NDR) to marshal the data. DCOM, since it is
built on top of RPC, as shown in FIGS. 3A & 3B, can leverage
RPC's functionality, such as the use of the NDR standard format. As
is known by those of skill in the art, the NDR standard is a
"receiver makes right" standard. Should any translation of the data
be necessary for a client computer and a server computer to
communicate, it is the receiver's responsibility to convert the
data into the format it requires. Thus, it is the receiver which
makes the data format "right" for itself. The sender, therefore,
need not perform any conversion on the data.
[0045] With reference again to FIGS. 3A & 3B, the marshalling
layer 122, in FIG. 3A, using the NDR standard, marshals data 144 by
reading parameters, such as pointers to arrays, or pointers to
integers, placed by the DCOM client 120 onto the memory stack 121
of the client computer 70. As is known by those skilled in the art,
a call to a function passes parameters including immediate data
values, such as integers or floating-point numbers, and pointers to
additional data, such as pointers to arrays of data, pointers to
text strings, or pointers to complex data structures. Therefore,
when the DCOM client 120 makes a call to the DCOM server 142, it
places onto the stack 121 a parameter set 125 include immediate
data and pointers for the current call. The pointers in the
parameter set 125 on the stack 121 point to the data 144 which is
in the client computer memory 145. The marshalling layer 122 first
determines the required size of the RPC buffer for holding all
marshaled data and the DCOM header, and requests such a buffer 123.
Then it copies any immediate data in the parameter set 125 into the
buffer. It also traverses all of the pointers in the parameter set
125 to retrieve all the data 144 and copies them into the RPC
buffer so that the call can be properly unmarshaled by the
marshaling layer 140, in FIG. 3B. The DCOM run-time layer 124 adds
the DCOM header and passes the buffer to the RPC run-time layer,
requesting that the data be sent across the network. As will be
known by one skilled in the art, an additional copy 129 may be made
by the operating system and placed into a protected-mode socket
buffer 131 for transmission across network connection 90. At the
server side, shown in FIG. 3B, the transmitted data 135 may be
delivered into a protected-mode socket buffer 133 and from there
into the RPC buffer 141. The marshaling layer 140 unmarshals the
marshaled data 137 from RPC buffer 141 into the parameter set 139
and the server memory 147 on the server computer 72. After the
pointers in the parameter set 139 on the stack 143 have been
recreated, the call made by the DCOM client 120 can then be made by
the stub 140 to the DCOM server 142.
[0046] As described above, to perform marshaling, the proxy 122
copies immediate data from the parameter set 125 and additional
data 144 to an RPC buffer 123 for transmission. This buffer holds a
duplicate 127 of the parameter set 125 and data, since the
parameter set is still resident in the stack 121 and the data is
still resident in the memory 144 from which it was copied. The
present invention contemplates a direct marshaling which stores
only a list of pointers to the memory locations of the data (either
in the parameter set 125 or additional data 144) in the RPC buffer,
rather than duplicating the data itself. As is known by those of
skill in the art, a pointer is generally of much smaller memory
size than the data it points to, especially in the case of pointers
to large data arrays. Therefore, the speed at which the proxy 122
and the stub 140 could marshal and unmarshal the parameters would
be increased by copying only pointers into the RPC buffer.
[0047] Certain network interface cards (NICs) provide for the
ability to perform scatter-gather operations. One example of a
network which can use the scatter-gather functionality of such a
NIC is a network based on the Virtual Interface Architecture (VIA),
which is an instance of a SAN. VIA allows network user-mode
applications to directly request a scatter-gather mode of
transmission without incurring the extra cost of a copy into a
protected-mode buffer. The gathering and the sending of the data is
performed directly by the NIC and requires no intervention by the
host processor, thereby eliminating the kernel-mode protocol stacks
130 and 132. Furthermore, the request to send and receive data is
presented by the user-mode application directly to the NIC without
transitioning into the operating system kernel and associated
device drivers.
[0048] In accordance with the invention, and turning to FIGS. 4A
& 4B, a modified proxy 198 marshals a list of pointers 125 into
the RPC buffer 123. An RPC flag can be added to the modified proxy
198 and stub 199, described in more detail below, to allow them to
inform the RPC run-time layers 126 and 136 that the RPC buffers 123
and 141 contain only a list of pointers to memory. The DCOM runtime
layers 124 and 138, as described above, can use the RPC layers to
communicate the call parameters across the network. Therefore, the
DCOM layers only pass the flag indicating that direct marshalling
is used, after adding or removing the DCOM header to or from the
packet, in a manner analogous to that of the known system described
above. The RPC runtime layers 126 and 136, however, can be modified
to accept a larger data structure from the modified proxy 198 and
stub 199. The larger data structure can accommodate the flag set by
the modified proxy 198 and stub 199. If the flag is set, the RPC
run-time layers 126 and 136 interpret the data 152 and 156 as a
list of scatter-gather entries, each comprising a starting memory
address of the data they point to and the length of the data. As
shown in FIG. 4A, the RPC run-time layer 126 adds RPC headers to
the list 152 and passes it to the loadable transport layer 128. The
loadable transport layer 128 then passes the list to the user mode
stack 148, which implements the implicit flow control, described in
more detail below. The user mode stack 148 stores the list 153 in a
list buffer 151. The NIC, at the VIA network connection layer 150,
gathers the immediate data from the parameter set 125 and
additional data 144, both pointed to by the list 153, out of the
stack 121 and the client memory 145 and transmits it across the
network.
[0049] On the receiving side, shown in FIG. 4B, since the
servers-side RPC runtime may receive calls on any methods supported
by the server process, it is in general not possible to specify a
receive scatter list for any arbitrary method in advance. However,
since the receiving RPC buffer is dedicated to the on-going RPC
call for its entire duration, the stub code and the server object
can use the data in the RPC buffer directly without first copying
the data into the data 146 in server memory 147, unless data format
conversion is required.
[0050] Turning to FIGS. 5A & 5B, a response from the DCOM
server object 142 to the DCOM client 120 is shown. As shown in FIG.
5B, a modified stub 199 marshals a list of pointers 157 into an RPC
buffer 141. The list of pointers 157 contains entries for any
immediate return data in the return parameter set 139 and any
additional data 146 in server memory 147. An RPC flag can be added
to the modified proxy 198 and stub 199, as described above, to
allow them to inform the RPC run-time layers 126 and 136 that the
RPC buffers 123 and 141 contain only a list of pointers to memory.
In a manner analogous to that described above in reference to FIGS.
4A & 4B, the DCOM run-time layer 138 can add a DCOM header to
the list. The RPC run-time layer 136 can then add RPC headers to
the list 157 and pass it to the loadable transport layer 134. The
loadable transport layer 134 then passes the list to the user mode
stack 149, which implements the implicit flow control, described in
more detail below. The user mode stack 149 stores the list 158 in a
list buffer 153. The NIC, at the VIA network connection layer 150,
gathers all the data pointed to by the list 158 and transmits it
across the network.
[0051] Because the NIC performs the gather and send operation
directly from the server memory 147, the DCOM server 142 should
delay clearing the memory 147 until after the NIC has finished
gathering and transmitting the data. If the DCOM server object 142
were to clear the memory prior to the completion of the send
operation, the NIC would not be able to gather the appropriate data
146 pointed to by the pointers 158. To insure that the DCOM server
object 142 does not prematurely clear data 146 but can still
reclaim the memory after its usage, a callback function can be
implemented at the marshaling layer 199 to be invoked by the
loadable transport layer 134. Initially, when the modified stub 199
intends to use the direct marshalling, it passes down to the
loadable transport layer 134 a context pointer and a function
pointer to the callback function. The loadable transport layer 134
then calls the callback function by supplying the context pointer
as a parameter to indicate that the sending of the data 146 has
completed.
[0052] On the receiving side, shown in FIG. 5A, if the client knows
the size of each piece of the returning data when it makes the
call, it can pass down a list of pointers for scattering in 159. In
this case, when the NIC receives the transmitted data, it can
scatter it directly into data 144 in client memory 145, as
instructed by the list of pointers 159. If the size of the
returning data cannot be determined beforehand, the incoming data
needs to be received by the RPC buffer 123 first. Unlike the server
side, it is undesirable for the client to use the data from buffer
123 directly because the client may need to hold on to the data
beyond the end of current call. Therefore, the data 160 in buffer
123 needs to be copied to data 144 in client memory 145 so that the
RPC buffer 123 can be released.
[0053] The proxy 78 and stub 80 shown in FIG. 2 implement standard
marshaling. Should a user choose to do so, they could write a
modified proxy 198 and stub 199 to perform the direct marshaling of
the present invention, as disclosed above. One method for doing so
is to manually replace the copy marshalling code with code which
constructs a scatter-gather list and to move the buffer release
code into an added callback function. The pointer to the buffer,
the context pointer, and the pointer to the callback function, as
described above, are passed to the RPC run-time layer and loadable
transport layers. Those layers, as also described above, can then
invoke the callback function by supplying the context pointer as a
parameter when the buffer can be released. An alternative method
for generating a modified proxy 198 and stub 199 to perform the
direct marshaling of the present invention would be to use an IDL
compiler. An interface can be described by the Interface Definition
Language (IDL). As is known by those skilled in the art, once an
interface is defined in IDL, an IDL compiler can create the code
for both a proxy and a stub capable of marshaling the defined
interface. One such IDL compiler is the Microsoft IDL (MIDL)
compiler from Microsoft Corporation. The IDL compiler could be
modified to automatically produce a proxy and stub capable of
providing such direct marshaling. Alternatively, a command-line
flag could be provided to indicate that direct marshalling is
requested.
[0054] As can be seen, the present invention reduces the number of
copies on the client side when calling an interface by marshaling
only a list of pointers into the RPC buffer 123 and allowing the
network to access the data directly. On the server side, the
received data can be retained in the RPC buffer 141, without
copying to memory 147. Additionally, such lists of pointers can be
used to eliminate a copy on the server side when sending a response
from the DCOM server 142, and when receiving the response on the
client side. The elimination of these copies results in a more
efficient transfer process from the DCOM client 120 to the DCOM
server 142 and from the server to the client.
[0055] An additional optimization contemplated by the present
invention is the removal of dispatching redundancies between the
DCOM run-time layers 124 and 138 and the RPC run-time layers 126
and 136. As is known by those skilled in the art, dispatching is
the process by which the called interface is located by the server
computer 72. Generally dispatching identifies an interface with
succeeding levels of specificity. Turning to FIG. 6, RPC
dispatching and DCOM dispatching on the server computer 72 are
illustrated. In order for a call from an RPC client to arrive at
the correct server computer 72, the call can specify the Internet
Protocol (IP) address of the server to which it is making the call.
Each server can then have multiple port address on which a call can
be made, and a port address can be specified in the call. The call
can also specify the interface identifier (IID) of the RPC
interface to which the call is made. Multiple interfaces can be
called through a single port. Thus, as shown in FIG. 6, the server
72 first checks the IP address at step 170. If the IP address
corresponds to the server 72, the server checks which port address
is specified in the call at step 168. Once the port address is
determined, the server can pass the call to the appropriate RPC
dispatching at step 166 which, depending on whether IID1 or IID2
was called can direct the call to interface IID1 at step 162 or
interface IID2 at step 164.
[0056] DCOM dispatching is performed in a similar manner, since
DCOM is layered on top of RPC, as can be seen from FIGS. 3A, 3B,
4A, 4B, 5A, and 5B. As shown in FIG. 6, the server 72 first checks
the IP address at step 184. If the IP address corresponds to the
server 72, the server checks which port address is specified in the
call at step 182. Once the port address is determined, the server
can pass the call to the appropriate DCOM dispatching element at
step 180. Unlike RPC, however, a single DCOM object, such as object
76 in FIG. 2 can have multiple interfaces, such as interface 82.
Because each interface has a unique interface identifier (IID) only
within the object providing that interface, it is possible for two
interfaces of two separate DCOM objects to have an identical IIDs.
Furthermore, because multiple DCOM objects can be called through a
single port, it is possible for two interfaces of two separate DCOM
objects, each of which can be called through the same port, to have
an identical IIDs. Therefore, to uniquely identify an interface in
such an environment, DCOM can use an interface pointer identifier
(IPID). The IPID is a combination of the IID of the interface, and
the object identifier (OID) of the object providing the interface.
In such a manner, the IPID can uniquely identify the interface by
referencing both the IID and the OID.
[0057] Therefore, in FIG. 6, when the server 72 passes the call to
the DCOM dispatching element at step 180, the DCOM dispatching
element determines the appropriate IPID to which the call is
directed. The DCOM dispatching element at step 180 is also known as
the stub manager, because it directs the call to the appropriate
stub, such as stub 80 in FIG. 2, based on the IPID. Thus, in FIG.
6, if the call was directed to interface IPID1, the stub manager at
step 180 can pass the call to stub1 at step 176 to call the
interface. Alternatively, if the call was directed to interface
IPID2, the stub manager at step 180 can pass the call to stub2 at
step 178 to call the interface. As is known by those skilled in the
art, a single stub can access multiple interfaces. Thus, it is not
required that the stub manager at step 180 invoke a different stub
for each call to a different interface.
[0058] As can be seen, DCOM relies on an IPID, a combination of an
IID and an OID to perform dispatching, while RPC relies only on the
IID. Nevertheless, because of the similarities between the two
systems, DCOM can be implemented to take advantage of RPC
dispatching. The RPC run-time layer 136 implements an RPC
dispatching 186 layer, as shown in FIG. 6, and an additional layer
188 to handle further RPC duties. The RPC run-time layer 126 does
not implement the dispatching layer 186, as a client computer does
not dispatch a call. RPC layer 188 can include RPC security, RPC
thread management, RPC socket connection management, and RPC
association management. The DCOM run-time layer 138, which
implements the DCOM dispatching 190, can be thought of as built on
top of the RPC dispatching 186 and the RPC layer 188, as shown in
FIG. 6. The DCOM run-time layer 124 on the client computer 70 does
not implement dispatching, but can provide DCOM security. To ensure
that the RPC run-time layer 126 and 136 is ready to send and
receive calls, the DCOM run-time layer 124 can still specify an IID
to which the call is directed. In such a way the RPC dispatching
186 on the server 72 can handle a DCOM call as it would an RPC
call. However, the DCOM dispatching 190 can use an IPID to uniquely
specify the interface to which the call is directed. The RPC IID is
therefore redundant, and used solely for the purpose of utilizing
the RPC layer 188. Furthermore, when the RPC dispatching 186
receives the IID, it performs additional functions, such as setting
up a communication dictionary, which are redundant in light of the
DCOM dispatching 190. The redundancy is significant because the
functions performed to enable RPC dispatching sometimes require a
network communication from the client computer 70 to the server
computer 72 and a return communication from the server to the
client. As is known by those skilled in the art, communications
across a network connection require significantly more time than
communications local to the client computer 70.
[0059] The present invention, therefore, contemplates removing the
use of the RPC IID and the attendant inefficiencies while
maintaining the useful RPC security, thread management, socket
connection management, and association management. One method for
doing so is to remove the RPC dispatching 186 and allow the DCOM
run-time layer 124 to specify only an IPID. On the server computer
72, when the RPC run-time layer 136 completes the steps required by
the RPC utility layer 188, such as confirming the authenticity of
the client computer 70, it performs the RPC dispatching in layer
186. However, as described above, the processing of the RPC
dispatching layer 186 is inefficient, as the DCOM dispatching layer
190 will perform its own dispatching when the RPC dispatching is
finished. Furthermore, the RPC dispatching layer 186 will only
provide a pointer into the DCOM dispatching layer 190. Therefore,
the present invention contemplates providing the pointer to the
DCOM dispatching layer 190 directly to the RPC utility layer 188.
In such a manner, when the RPC utility layer 188 completes its
tasks, it can pass the call directly to the DCOM dispatching layer
190.
[0060] An additional modification which can be made as a result of
the change to the RPC run-time layer 136 described above, is to
remove those API calls to the RPC dispatching 186 from the DCOM
run-time layer 138. An example of the APIs affected is shown in
Table 1 below. As can be seen from the table, the API calls that
can be removed are those that perform functions attendant with RPC
dispatching.
1 TABLE 1 RPC_STATUS RPC_ENTRY RpcServerInqIf( RPC_STATUS RPC_ENTRY
RpcServerRegisterIf( RPC_STATUS RPC_ENTRY RpcServerRegisterIfEx(
RPC_STATUS RPC_ENTRY RpcServerUnregisterIf( RPC_STATUS RPC_ENTRY
RpcServerUseAllProtseqsIf( RPC_STATUS RPC_ENTRY
RpcServerUseAllProtseqsIfEx( RPC_STATUS RPC_ENTRY
RpcServerUseProtseqIf( RPC_STATUS RPC_ENTRY
RpcServerUseProtseqIfEx(
[0061] On the client computer 70, the DCOM run-time layer 124 can
be modified by removing the code that sends the IID together with
the IPID. Because the server 72 no longer performs RPC dispatching
prior to DCOM dispatching, there is no longer a need for the IID.
As was described above, the IPID uniquely identifies the interface
called, and is (conceptually) composed of a combination of the IID
and the OID. Therefore, efficiencies are achieved due to the
elimination of repeated tasks between the DCOM run-time layer 138
the RPC run-time layer 136 on the server machine; most notably the
RPC run-time layer 136 no longer performs a round-trip network
communication when the call is first initiated.
[0062] Another improvement to the speed at which DCOM runs over a
network can be achieved through a modification of the RPC flow
control. Flow control ensures that each packet sent across a
network is being expected by the intended recipient. FIG. 7
illustrates a known flow control scheme, which uses an explicit
flow control. With such an explicit flow control, the sender must
wait until the receiver signals it is ready to receive prior to
sending a packet across the network. Thus, in FIG. 7, the client
computer 70 waits at step 204 for the server computer 72 to
indicate that it is ready to receive. An OK TO SEND message can be
one such indication that the computer is ready to receive. At step
200 the server computer 72 can post a receive buffer. By posting
the receive buffer, the server computer 72 is setting aside memory
in which to receive whatever data the client computer 70 will send.
Meanwhile, at step 204, the client computer 70 is idling, waiting
for permission to send the data. Once the server computer 72 has
posted the receive buffer at step 200, it can send an OK TO SEND
message 202. When the client computer 70 receives the OK TO SEND
message 202, it can then proceed, at step 206 to send its request
to the server computer 72. The request is sent as data 208, which
is received by the server computer and placed in the receive buffer
at step 210. Once the request has been received, the server can
perform whatever work is required by the request at step 212. When
the server computer 72 has finished the work which was requested by
the client, it idles, at step 214, waiting for the client to signal
that it is ready to receive. The client computer can post its
receive buffer at step 216, and can then send an OK TO SEND message
218. When the server receives the OK TO SEND message 218, it can
send its response to the client's request at step 220. The response
data 222 is received by the client and can be placed in the receive
buffer at step 224.
[0063] As can be seen from FIG. 7, the known flow control protocol
results in two OK TO SEND messages for each client request and
server response. The flow control messages thus account for 50% of
the message traffic. Furthermore, the efficiency of the system is
reduced when the client and server computers wait for one another
to send explicit flow control messages, not performing any useful
work in the interim. The present invention contemplates removing
the overhead and the inefficiency of the known system by
implementing an implicit flow control at the loadable transport
layers 128 and 134 and turning off the explicit flow control in the
user mode stack 148 and 149. An implicit flow control relies on RPC
semantics to ensure that each packet sent has a destination that is
capable of receiving it. The implicit flow control of the present
invention is shown in FIG. 8. As can be seen, neither of the OK TO
SEND explicit flow control transmissions of FIG. 7 are present. The
explicit flow control messages can be eliminated because, prior to
sending any data, each computer can first pre-post a receive
buffer. Therefore, the very act of sending a message is an
indication to the other computer that the receive buffer is already
posted and the computer is ready to receive. Thus, the other
computer need not wait for an explicit flow control message; rather
the receipt of any message is an implicit OK TO SEND. This implicit
flow control, as will be described in more detail with reference to
FIG. 8, allows for the removal of the flow control messages.
Furthermore, as will be described in more detail with reference to
FIG. 9, the implicit flow control of the present invention can
increase the efficiency of the overall system by reducing the
unproductive computer time spent waiting for explicit flow control
messages to be sent.
[0064] The implicit flow control of the present invention requires
the size of the pre-posted buffer to be sufficiently large so that
it may accept whatever data was sent. Generally a default size can
be chosen, such as the Maximum Transfer Unit (MTU) of the network.
An overflow of the pre-posted buffer will result in an error and a
request to resend the data, delaying processing of the data. As an
alternative, a default size smaller than the MTU could be chosen,
decreasing the amount of resources used, but increasing the number
of overflows and resend requests, and thereby decreasing the
overall speed of the system.
[0065] In an RPC context, there can exist non-RPC communication. A
non-RPC communication is a communication in which at least one of
the two parties has no receive operation following its send
operation. In such a case, the optimized flow control may not be
applicable, because the total number of messages is not an even
number. One example of such a non-RPC communication is an
authentication between a client and a server. Authentication can be
of the form: request by client, challenge with counter-challenge by
server, and response by client. This sequence of three messages can
be made to maintain RPC semantics by simply adding a fourth message
back to the client from the server. Alternatively, the client's
first method call can be combined with the third authentication
message, the response by the client, so that one message
accomplishes both functions, and an even number of messages is
maintained.
[0066] FIG. 8 illustrates a steady-state situation in the implicit
flow control case. Initially, the server computer 72 can post a
receive buffer so that it may be able to receive a message from the
client computer 70. In FIG. 8, it is assumed that the receive
buffer used at step 236 was posted in a prior sequence. Therefore,
beginning with step 230, the client computer 70 pre-posts a receive
buffer. Once the client computer 70 has posted the receive buffer,
it can send its request to the server. Thus, at step 232, the
client sends data 234 to the server. The server, at step 236,
places the data 234 into the receive buffer that was posted during
an earlier cycle, as described above. At step 238, the server
computer 72 can perform the work requested by the client. At step
240, prior to sending the response to the client, the server can
pre-post a receive buffer. Once the receive buffer is posted, at
step 242 the server computer 72 can send to the client computer 70
data 244, which can be placed by the client in the receive buffer
at step 246. The receive buffer into which the response is placed
at step 246 is the receive buffer which was posted at step 230. The
sending of data 234 to the server was an implicit OK TO SEND from
the client to the server. Thus, at step 242, the server could send
the response data 244 without waiting for an explicit OK TO
SEND.
[0067] As explained above, an explicit flow control system reduces
the efficiency of the overall system by requiring each computer to
idle while waiting for an explicit OK TO SEND flow control message.
The implicit flow control of the present invention can improve the
system efficiency by reducing the amount of time each computer
idles waiting for explicit flow control messages. As can be seen
with reference to FIG. 8, when the server computer 72 has finished
its computations at step 238, it is ready to send the response data
244 to the client computer 70. The server 72 need not wait for an
explicit flow control message. Rather, it can send data 244 as soon
as it has pre-posted the receive buffer 240. Thus, the amount of
time between the completion of work at step 238 and the sending of
the response at step 242 is minimized. The client computer 70,
therefore, receives a result from the server 72 in less time than
in the known explicit flow control environment.
[0068] As will be known by those skilled in the art, in addition to
posting receive buffers, each computer can clear those buffers once
the data stored therein is no longer needed by the computer. In
such a manner the same physical memory of a computer can be reused,
ensuring that communication can continue indefinitely. Furthermore,
as is also known by those of skill in the art, the sending of data
from one computer to another, such as the send operations in steps
232 and 242 of FIG. 8, requires that the data 234 and 244 be placed
in a send buffer for transmission. Therefore, there also exist send
buffers which can also be cleared, in a manner analogous to the
clearing of the receive buffers.
[0069] In the known explicit flow control, the client and the
server could clear the buffers while waiting for an explicit OK TO
SEND message. The present invention, however, provides greater
flexibility with respect to the timing of the clear send buffer and
clear receive buffer operations. Those operations can be delayed so
that a computer can receive a request, perform work, and return a
result in the least amount of time, thereby increasing the overall
efficiency of the system. FIG. 9 illustrates the sequence of the
buffer clearing operations, as contemplated by the present
invention. As with FIG. 8, FIG. 9 illustrates a steady-state
system. The index shown in FIG. 9 indicates the round of
communication between the client and server computers. Because FIG.
9 illustrates a steady-state system, which can exist at any time,
the index n is used to indicate any integer. Thus, the receive
buffer posted at step 250 will receive the nth response from the
server computer. As will be described later, at step 268 the client
computer pre-posts a receive buffer which will receive the next
response from the server, or the n+1th response.
[0070] As described in detail above, the implicit flow control of
the present invention contemplates that the receive buffers can be
posted prior to the sending of any data, so that the sending of the
data itself is the implicit OK TO SEND message to the other
computer. Thus, prior to sending the nth request at step 252, the
client computer 70 pre-posts the receive buffer at step 250 to
accept the response to the nth request: the nth response. At step
251 the client 70 performs the work which necessitates the request
to the server 72. The client can then send the request at step 252
as data 254. At step 256, the server computer 72 receives the data
254 into a receive buffer which was posted during an earlier cycle,
not shown. As can be seen from the figure, prior to the receipt of
the nth request from the client at step 256, the server computer
cleared, at step 253, the receive buffer from the request prior to
the nth request, the n-1th request. Similarly, at step 255, the
server computer 72 cleared the send buffer from the n-1th reply.
Therefore, the buffers which are cleared are those from the round
of communication prior to the current round.
[0071] At step 258, the server does work on the request, and at
step 260, the server pre-posts the receive buffer for the coming
request, the n+1th request. At step 262 the server sends the
response to the request of the current round, the nth round, and
that is received by the client at step 266 into the receive buffer
which was posted at step 250. While the server computer 72 was
performing work at step 258, the client computer 70 cleared the
receive buffer from the prior round, the n-1th round in step 261
and cleared the send buffer from the prior round in step 263, as
shown in FIG. 9. Similarly, after the server 72 sent the response
at step 262, it cleared the receive buffer at step 267 and the send
buffer at step 274. The send buffer cleared in step 274 was the
buffer used to send the response in step 262.
[0072] The cycle of pre-post receive buffer, perform work, send
request or reply, clean up receive, clean up send, and receive
reply or request repeats itself for each request and response
cycle. A second cycle is illustrated in FIG. 9, as steps 268
through 285. The second cycle reveals that the clean up operations
are interleaved with the send and receive operations, and trail by
one round. For example, the buffer pre-posted at step 250 by the
client receives data at step 266 and is cleared at step 279.
Similarly, the send buffer used by the client at step 252 is
cleaned up at step 281. On the server side, the buffer into which
data was received at step 256 is cleared in step 267. Also on the
server, the receive buffer that was pre-posted in step 260 received
data in step 276, and the send operation in step 262 was cleaned up
in step 274.
[0073] Because the implicit flow control of the present invention
allows the clean up operations to be flexibly timed, they can be
scheduled to take place during the time in which the computer is
waiting for a response. For example, when the client computer 70
sends the request at step 252, it may not be able to do substantive
work until it receives a reply at step 266. Therefore, scheduling
buffer clearing operations, such as those at steps 261 and 263 in
the time between step 252 and 264 provides for efficiencies at the
client side. At the server side, the server 72 can decrease the
time the client may be waiting for a response by decreasing the
time between step 256 when it receives a request, and step 262 when
it sends the response. Thus, in the interim, the server 72 only
performs the work requested by the client, at step 258, and
preposts the receive buffer at step 260, as contemplated by the
implicit flow control of the present invention. The server 72 can
then clear its buffers after step 262 when the client may no longer
be waiting for it to respond to the client's request. As can be
seen, by scheduling the clearing of the send and receive buffers
outside of the receive operation and the responding send operation
cycle, the overall efficiency of the system can be increased by
decreasing the time each computer waits for the other to
respond.
[0074] The present invention provides three mechanisms for
enhancing the speed of DCOM over a network. At the application
level, copying is reduced by using an array of pointers rather than
the values themselves and taking advantage of the network interface
card's ability to do scatter-gather. The duplication of effort
between the RPC runtime and the DCOM runtime is eliminated by
removing the dispatching of the RPC runtime layer. Finally, at the
transport level, the flow control is accelerated by switching to an
implicit flow control, and scheduling the clear buffers commands
outside of the critical time to reduce idle. Each of the three
mechanisms can be used by themselves or in any combination to
achieve a speed increase over the prior art.
[0075] All of the references cited herein, including patents,
patent applications, and publications, are hereby incorporated in
their entireties by reference.
[0076] In view of the many possible embodiments to which the
principles of this invention may be applied, it should be
recognized that the embodiment described herein with respect to the
drawing figures is meant to be illustrative only and should not be
taken as limiting the scope of invention. For example, those of
skill in the art will recognize that the elements of the
illustrated embodiment shown in software may be implemented in
hardware and vice versa or that the illustrated embodiment can be
modified in arrangement and detail without departing from the
spirit of the invention. Therefore, the invention as described
herein contemplates all such embodiments as may come within the
scope of the following claims and equivalents thereof.
* * * * *