U.S. patent application number 09/169223 was filed with the patent office on 2002-02-28 for load balancing cooperating cache servers by shifting forwarded request.
Invention is credited to JORDAN, KEVIN MICHAEL, WU, KUN-LUNG, YU, PHILIP SHI-LUNG.
Application Number | 20020026560 09/169223 |
Document ID | / |
Family ID | 22614704 |
Filed Date | 2002-02-28 |
United States Patent
Application |
20020026560 |
Kind Code |
A1 |
JORDAN, KEVIN MICHAEL ; et
al. |
February 28, 2002 |
LOAD BALANCING COOPERATING CACHE SERVERS BY SHIFTING FORWARDED
REQUEST
Abstract
In a system including a collection of cooperating cache servers,
such as proxy cache servers, a request can be forwarded to a
cooperating cache server if the requested object cannot be found
locally. An overload condition is detected if for example, due to
reference skew, some objects are in high demand by all the clients
and the cache servers that contain those hot objects become
overloaded due to forwarded requests. In response, the load is
balanced by shifting some or all of the forwarded requests from an
overloaded cache server to a less loaded one. Both centralized and
distributed load balancing environments are described.
Inventors: |
JORDAN, KEVIN MICHAEL;
(BRIARCLIFF, NY) ; WU, KUN-LUNG; (YORKTOWN
HEIGHTS, NY) ; YU, PHILIP SHI-LUNG; (CHAPPAQUA,
NY) |
Correspondence
Address: |
KEVIN M JORDAN
IBM CORPORATION
INTELLECTUAL PROPERTY LAW DEPARTMENT
PO BOX 218
YORKTOWN HEIGHTS
NY
10598
|
Family ID: |
22614704 |
Appl. No.: |
09/169223 |
Filed: |
October 9, 1998 |
Current U.S.
Class: |
711/120 ;
709/203; 711/130; 711/144; 711/147 |
Current CPC
Class: |
G06F 9/505 20130101 |
Class at
Publication: |
711/120 ;
711/130; 709/203; 711/144; 711/147 |
International
Class: |
G06F 012/08 |
Claims
We claim:
1. A cache server load balancing method, comprising the steps of:
receiving forwarded requests from a cooperating cache server in
response to a cache miss for an object on the cooperating cache
server; and shifting one or more of said forwarded requests for the
object between cooperating cache servers based on a load condition
and a forwarding frequency for the object.
2. The method of claim 1, said shifting step further comprising the
steps of: periodically monitoring the load condition on and the
forwarding frequency to an owning cache server; and proactively
shifting one or more subsequent forwarded requests for the cached
object from the owning cache server to one or more of said
cooperating cache servers, in response to said monitoring.
3. The method of claim 1, said shifting step further comprising the
step of checking the load condition and forwarding frequency, in
response to the forwarded request.
4. The method of claim 1, wherein said shifting comprises the step
of modifying an ownership for the object to a shared ownership
between two or more of said cooperating cache servers.
5. The method of claim 4, further comprising the step of merging
said shared ownership in response to change in the load
condition.
6. The method of claim 1, further comprising the step of locally
monitoring the load on each cooperating cache server.
7. The method of claim 6, further comprising the step of: a
distributed load monitor monitoring and maintaining a local load
condition, the forwarding frequency and ownership information for
cached objects on said each cooperating cache server.
8. The method of claim 7, further comprising the steps of: said
cooperating cache servers periodically exchanging and maintaining
one or more of: the load condition information; the forwarding
frequency; and the ownership information.
9. The method of claim 7, further comprising the steps of: said
cooperating cache servers exchanging by piggybacking one or more
of. the load condition information; the forwarding frequency; and
the ownership information; with one or more of the forwarded
requests and responses.
10. The method of claim 1, further comprising the step of.
receiving a forwarded request and updating the forwarding
frequency.
11. The method of claim 7, further comprising the steps of:
identifying a less loaded cooperating cache server; and
communicating one or more of: a shift request; and a copy of the
cached object, to said less loaded cooperating cache server.
12. The method of claim 11, further comprising the steps of: said
less loaded cooperating cache server receiving said shift request;
and said less loaded cooperating cache server requesting a copy of
the object from an originating object server, in response to said
shift request.
13. The method of claim 11, wherein the copy is obtained via one or
more of an intranet, WAN or Internet.
14. The method of claim 1, further comprising the step of multi
casting a shift request message to one or more of the other
cooperating cache servers so that subsequent forward requests will
be shifted.
15. The method of claim 14, further comprising the step of: the
cooperating cache servers maintaining one of a local copy of a
caching table and modifying a hash function; and the cooperating
cache servers modifying the ownership information by one of.
updating a local copy of a caching table; and modifying a hash
function.
16. The method of claim 15, further comprising the steps of:
modifying the ownership for the object to a shared ownership
between at least two of said cooperating cache servers; and said
cooperating cache servers forwarding subsequent object requests to
one or more less loaded shared owners of the object.
17. The method of claim 16, further comprising the steps of:
detecting a decrease in the load condition for a shared object; and
merging the shared ownership, in response to the decrease in the
load condition.
18. The method of claim 1, wherein said shifting one or more of
said forwarded requests comprises the steps of: communicating a
copy of the object from an owning cache server to one or more of
said cooperating cache servers; and said cooperating cache server
receiving and caching the copy of the object.
19. The method of claim 1, further comprising the steps of,
calculating the load condition of each cache server in past
intervals; computing a mean load of all cache servers in past
intervals; and finding the cache servers that exceed a threshold
above said mean load.
20. The method of claim 1, wherein the load condition of said
cooperating cache server can be a weighted sum of a count of said
forwarded requests, and a count of direct requests to said
cooperating cache server.
21. The method of claim 1, further comprising the step of
maintaining cache information at one or more of: each object level;
and a partition of objects level.
22. The method of claim 21, wherein said cache information of said
object level or said partition comprises the forwarding frequency
associated with the object.
23. The method of claim 22, further comprising the step of: a
distributed load monitor monitoring and locally maintaining load
conditions, forwarding frequency and ownership information for
cached objects on each cache server.
24. The method of claim 23, further comprising the steps of. said
cooperating cache servers periodically exchanging one or more of
the load condition, the forwarding frequency and the ownership
information.
25. The method of claim 22, further comprising the steps of: said
cooperating cache servers exchanging by piggybacking one or more
of. the load condition; the forwarding frequency; and the ownership
information; with one or more of the forwarded requests and
responses.
26. In a collection of cooperating cache servers, where each cache
server can receive direct requests and forwarded requests, and upon
a cache miss, a request can be forwarded to an owning cache server
caching said object, a load balancing method comprising the steps
of, monitoring a load condition and a forwarding frequency for said
cooperating cache servers; and shifting one or more forwarded
requests from one cooperating cache server to a second cooperating
cache server based on a change in the load condition and the
forwarding frequency.
27. The method of claim 26, wherein said step of monitoring the
load condition comprises the steps of, calculating the load
condition of each cache server in past intervals; computing a mean
load of all proxy cache servers in past intervals; and finding
those proxy cache servers that exceed a threshold above said mean
load.
28. The method of claim 26, wherein said shifting step can be
performed in response to one or more of: said forwarded requests
from said cooperating cache servers; and periodically monitoring
the load condition and the forwarding frequency.
29. The method of claim 26, further comprising the step of a
centralized logical load monitor maintaining the forwarding
frequency and the load condition for the cooperating cache
servers.
30. The method of claim 26, wherein the load condition of said
cache server can be a weighted sum of: a count of forwarded
requests; and a count of direct requests to said cache server.
31. The method of claim 26, further comprising the step of
maintaining cache information at each object level or at a
partition of objects level.
32. The method of claim 31, wherein said cache information of the
object level or the partition level comprises the forwarding
frequency of requests through said load monitor to said object.
33. The method of claim 26, wherein said cooperating cache servers
comprise cooperating proxy cache servers.
34. The method of claim 26, further comprising the steps of: a
logical directory server maintaining a caching table and a load
table; said cache servers interrogating said directory server for
object locations in other cache servers for a locally missed
object; and said directory server load balancing requests among
said cache servers by manipulating said caching table, in response
to requests for object locations.
35. The method of claim 29, further comprising the steps of: each
cache server multicasting to a list of cooperating cache servers to
locate a copy of a locally missed object; and said shifting step
comprising the step of excluding overloaded cache servers from a
subset of neighboring cache servers for multicasting.
36. A program storage device readable by a machine, tangibly
embodying a program of instructions executable by the machine to
perform method steps for cache server load balancing, said method
steps comprising: receiving forwarded requests from a cooperating
cache server in response to a cache miss for an object on the
cooperating cache server; and shifting one or more of said
forwarded requests for the object between cooperating cache servers
based on a load condition and a forwarding frequency for the
object.
37. The program storage device of claim 36, said shifting step
further comprising the steps of: periodically monitoring the load
condition on and the forwarding frequency to an owning cache
server; and proactively shifting one or more subsequent forwarded
requests for the cached object from the owning cache server to one
or more of said cooperating cache servers, in response to said
monitoring.
38. The program storage device of claim 36, said shifting step
further comprising the step of checking the load condition and
forwarding frequency, in response to the forwarded request.
39. The program storage device of claim 36, wherein said shifting
comprises the step of modifying an ownership for the object to a
shared ownership between two or more of said cooperating cache
servers.
40. The program storage device of claim 39, further comprising the
step of merging said shared ownership in response to change in the
load condition.
41. The program storage device of claim 36, further comprising the
step of locally monitoring the load on each cooperating cache
server.
42. The program storage device of claim 41, further comprising the
step of: a distributed load monitor monitoring and maintaining a
local load condition, the forwarding frequency and ownership
information for cached objects on said each cooperating cache
server.
43. The program storage device of claim 42, further comprising the
steps of: said cooperating cache servers periodically exchanging
and maintaining one or more of: the load condition information; the
forwarding frequency; and the ownership information.
44. The program storage device of claim 42, further comprising the
steps of: said cooperating cache servers exchanging by piggybacking
one or more of: the load condition information; the forwarding
frequency; and the ownership information; with one or more of the
forwarded requests and responses.
45. The program storage device of claim 36, further comprising the
step of. receiving a forwarded request and updating the forwarding
frequency.
46. The program storage device of claim 42, further comprising the
steps of: identifying a less loaded cooperating cache server; and
communicating one or more of: a shift request; and a copy of the
cached object, to said less loaded cooperating cache server.
47. The program storage device of claim 46, further comprising the
steps of: said less loaded cooperating cache server receiving said
shift request; and said less loaded cooperating cache server
requesting a copy of the object from an originating object server,
in response to said shift request.
48. The program storage device of claim 46, wherein the copy is
obtained via one or more of an intranet, WAN or Internet.
49. The program storage device of claim 45, further comprising the
step of. multicasting a shift request message to one or more of the
other cooperating cache servers so that subsequent forward requests
will be shifted.
50. The program storage device of claim 49, further comprising the
step of: the cooperating cache servers maintaining one of: a local
copy of a caching table; and a hash function; and the cooperating
cache servers modifying the ownership information by one of:
updating the local copy of a caching table; and modifying a hash
function.
51. The program storage device of claim 50, further comprising the
steps of: modifying the ownership for the object to a shared
ownership between at least two of said cooperating cache servers;
and said cooperating cache servers forwarding subsequent object
requests to one or more less loaded shared owners of the
object.
52. The program storage device of claim 51, further comprising the
steps of: detecting a decrease in the load condition for a shared
object; and merging the shared ownership, in response to the
decrease in the load condition.
53. The program storage device of claim 36, wherein said shifting
one or more of said forwarded requests comprises the steps of:
communicating a copy of the object from an owning cache server to
one or more of said cooperating cache servers; and said cooperating
cache server receiving and caching the copy of the object.
54. The program storage device of claim 36, further comprising the
steps of, calculating the load condition of each cache server in
past intervals; computing a mean load of all cache servers in past
intervals; and finding the cache servers that exceed a threshold
above said mean load.
55. The program storage device of claim 36, wherein the load
condition of said cooperating cache server can be a weighted sum of
a count of said forwarded requests, and a count of direct requests
to said cooperating cache server.
56. The program storage device of claim 36, further comprising the
step of maintaining cache information at one or more of: each
object level; and a partition of objects level.
57. The program storage device of claim 56, wherein said cache
information of said object level or said partition comprises the
forwarding frequency associated with the object.
58. The program storage device of claim 57, further comprising the
step of: a distributed load monitor monitoring and locally
maintaining load conditions, forwarding frequency and ownership
information for cached objects on each cache server.
59. The program storage device of claim 58, further comprising the
steps of: said cooperating cache servers periodically exchanging
one or more of the load condition, the forwarding frequency and the
ownership information.
60. The program storage device of claim 57, further comprising the
steps of: said cooperating cache servers exchanging by piggybacking
one or more of: the load condition; the forwarding frequency; and
the ownership information; with one or more of the forwarded
requests and responses.
61. A program storage device readable by a machine, tangibly
embodying a program of instructions executable by the machine to
perform method steps for cache server load balancing in a
collection of cooperating cache servers, where each cache server
can receive direct requests and forwarded requests, and upon a
cache miss, a request can be forwarded to a cooperating cache
server caching said object, said method steps comprising:
monitoring a load condition and a forwarding frequency for said
cooperating cache servers; and shifting one or more forwarded
requests from one cooperating cache server to a second cooperating
cache server based on a change in the load condition and the
forwarding frequency.
62. The program storage device of claim 61, wherein said step of
monitoring the load condition comprises the steps of, calculating
the load condition of each cache server in past intervals;
computing a mean load of all proxy cache servers in past intervals;
and finding those proxy cache servers that exceed a threshold above
said mean load.
63. The program storage device of claim 61, wherein said shifting
step can be performed in response to one or more of: said forwarded
requests from said cooperating cache servers; and periodically
monitoring the load condition and the forwarding frequency.
64. The program storage device of claim 61, further comprising the
step of. a centralized logical load monitor maintaining the
forwarding frequency and the load condition for the cooperating
cache servers.
65. The program storage device of claim 61, wherein the load
condition of said cache server can be a weighted sum of: a count of
forwarded requests; and a count of direct requests to said cache
server.
66. The program storage device of claim 61, further comprising the
step of maintaining cache information at each object level or at a
partition of objects level.
67. The program storage device of claim 61, wherein said cache
information of the object level or the partition level comprises
the forwarding frequency of requests through said load monitor to
said object.
68. The program storage device of claim 61, wherein said
cooperating cache servers comprise cooperating proxy cache
servers.
69. The program storage device of claim 61, further comprising the
steps of: maintaining a caching table and a load table; receiving
requests for object locations in other cache servers for a locally
missed object; and load balancing requests among said cache servers
by manipulating said caching table, in response to said requests
for object locations.
70. The program storage device of claim 64, further comprising the
steps of: each cache server multicasting to a list of cooperating
cache servers to locate a copy of a locally missed object; and said
shifting step comprising the step of excluding overloaded cache
servers from a subset of neighboring cache servers for
multicasting.
71. A cache server load balancing system, comprising: means for
receiving forwarded requests from a cooperating cache server in
response to a cache miss for an object on the cooperating cache
server; and shifting means for shifting one or more of said
forwarded requests for the object between cooperating cache servers
based on a load condition and a forwarding frequency for the
object.
72. The system of claim 71, said shifting means further comprising:
means for periodically monitoring the load condition on and the
forwarding frequency to said cooperating cache servers; and means
for proactively shifting one or more subsequent forwarded requests
for the cached object from between said cooperating cache servers,
in response to said monitoring.
73. The system of claim 71, said shifting means further comprising
means for checking the load condition and forwarding frequency, in
response to the forwarded request.
74. The system of claim 71, wherein said shifting means comprises
means for modifying an ownership for the object to a shared
ownership between two or more of said cooperating cache
servers.
75. The system of claim 74, further comprising merging said shared
ownership in response to a change in the load condition.
76. The system of claim 71, further comprising means for locally
monitoring the load condition on each cooperating cache server.
77. The system of claim 76, further comprising: distributed load
monitor means for maintaining a local load condition, the
forwarding frequency and ownership information for cached objects
on said each cooperating cache server.
78. The system of claim 77, further comprising: means for said
cooperating cache servers to periodically exchange and maintain one
or more of: the load condition information; the forwarding
frequency; and the ownership information.
79. The system of claim 77, further comprising the steps of: means
for said cooperating cache servers to piggyback one or more of: the
load condition information; the forwarding frequency; and the
ownership information; with one or more of the forwarded requests
and responses.
80. The system of claim 71, further comprising: means for updating
the forwarding frequency, in response to a forwarded request.
81. The system of claim 77, further comprising: means for
identifying a less loaded cooperating cache server; and means for
communicating one or more of: a shift request; and a copy of the
cached object, to said less loaded cooperating cache server.
82. The system of claim 81, further comprising: said less loaded
cooperating cache server including means for receiving said shift
request; and said less loaded cooperating cache server including
means for requesting a copy of the object from an originating
object server, in response to said shift request.
83. The system of claim 8l, wherein the copy is obtained via one or
more of an intranet, WAN or Internet.
84. The system of claim 80, further comprising: multicasting means
for multi casting a shift request message to one or more of the
other cooperating cache servers so that subsequent forward requests
will be shifted.
85. The system of claim 77, further comprising: a local copy of a
caching table on said each cooperating cache server; and means for
said cooperating cache servers to maintain the forwarding frequency
and the ownership information based on the local copy of the
caching table.
86. The system of claim 77, further comprising: hash function means
for hashing the object space into a number of hash buckets much
larger than a total number of said cache servers and assigning said
hash buckets to said each cooperating cache server for locating a
copy of a locally missed object; and said shifting means comprising
means for moving one or more hash buckets from an overloaded server
to a less loaded server, effectively changing the hash function so
that forwarded requests will not go to said overloaded server.
87. The system of claim 86, further comprising: means for modifying
the ownership for the object to a shared ownership between at least
two of said cooperating cache servers; and means for said
cooperating cache servers to forward subsequent object requests to
one or more less loaded shared owners of the object.
88. The system of claim 87, further comprising: means for detecting
a decrease in the load condition for a shared object; and means for
merging the shared ownership, in response to the decrease in the
load condition.
89. The system of claim 71, wherein said shifting means for
shifting one or more of said forwarded requests comprises means for
communicating a copy of the object from an owning cache server to
one or more of said cooperating cache servers.
90. The system of claim 71, further comprising, means for
calculating the load condition of each cache server in past
intervals; means for computing a mean load of all cache servers in
past intervals; and means for finding the cache servers that exceed
a threshold above said mean load.
91. The system of claim 71, wherein the load condition of said
cooperating cache server comprises a weighted sum of a count of
said forwarded requests, and a count of direct requests to said
cooperating cache server.
92. The system of claim 71, further comprising means for
maintaining cache information at one or more of: each object level;
and a partition of objects level.
93. The system of claim 92, wherein said cache information of said
object level or said partition comprises the forwarding frequency
associated with the object.
94. The system of claim 71, further comprising: centralized logical
load monitor means for maintaining the forwarding frequency and the
load condition for said cooperating cache servers.
95. The system of claim 71, wherein said cooperating cache servers
comprise cooperating proxy cache servers.
96. The system of claim 98, further comprising: a logical directory
server means for locating objects and forwarding requests for
locally missed objects; means for maintaining a caching table and a
load table, coupled to said directory server; cache server means
for interrogating said directory server for object locations in
other cache servers for a locally missed object; and directory
server means for load balancing requests among said cache servers
by manipulating said caching table, in response to requests for
object locations.
97. The system of claim 71, further comprising: multicasting means
on each cache server for multicasting to a list of cooperating
cache servers to locate a copy of a locally missed object; said
shifting means comprising means for excluding overloaded cache
servers from a subset of neighboring cache servers for
multicasting.
Description
FIELD OF THE INVENTION
[0001] The present invention is related to load balancing among
cooperating cache servers and in particular to load balancing based
on load conditions and a frequency that requests are forwarded from
cooperating cache servers.
BACKGROUND
[0002] The growth in the usage of the World Wide Web has been
increasing exponentially. As a result, response times for accessing
web objects can become unsatisfactorily slow. One approach to
improving web access time is to employ one or more proxy cache
servers between browsers and the originating web servers. Examples
of proxy cache servers include a cluster of PC servers running
Microsoft's Windows NT.TM., such as the NETFINITY.TM. servers from
IBM; and workstation servers running IBM's AIX.TM. operating
system, such as the IBM RS/6000.TM. or SP/2.TM.. In fact, more and
more organizations, such as Internet Service Providers (ISPs) and
corporations, are using a collection of cooperating proxy cache
servers to help improve response time as well as reduce traffic to
the Internet. A collection of cooperating cache servers have
distinct advantages over a single cache server in terms of
reliability and performance. If one fails, requests can still be
serviced by other cooperating cache servers. Requests can be
distributed among the servers, thus increasing scalability.
Finally, the aggregate cache size is much larger so that it is more
likely that a requested object will be found in one of the cache
servers.
[0003] With cooperating cache servers, a request that cannot be
serviced locally due to a cache miss can be forwarded to another
cache server storing the requested object. As a result, there are
two kinds of requests that can come to a cache server: direct
requests and forwarded requests. Direct requests are those that are
received directly from clients. Forwarded requests are those that
come from other cooperating cache servers on behalf of their
clients due to cache misses on the cache servers. With requests
forwarded among the cache servers, a cache server can easily become
overloaded if it happens to contain in-demand (or"hot") objects
that most clients are currently interested in, creating uneven
workloads among the cache servers. Uneven workloads can create a
performance bottleneck, as many of the cache servers are waiting
for the same overloaded cache server to respond to requests
forwarded to it. Therefore, there is a need for a way to perform
dynamic load balancing among a collection of proxy cache servers.
The present invention addresses such a need.
[0004] Load balancing is traditionally done by a front-end
scheduler which"evenly distributes" incoming direct requests among
the cache servers. For example, load balancing can be done at the
DNS level by manipulating a mapping table, such as is done by the
NETRA.TM. proxy cache by Sun Microsystems ("Proxy Cache Server,
Product Overview", white paper, Sun Microsystems,
http://www.sun.com/). Load balancing among a cluster of servers can
also be done with a front-end router, such as the NETDISPATCHER.TM.
offered by IBM (see e.g., G. Goldszmidt and G. Hunt,"NetDispatcher:
A TCP Connection Router," IBM Research Report, RC 20853, May 1997).
Here, incoming requests are distributed by the NETDISPATCHER.TM. to
the least loaded server in the cluster. However, these traditional
approaches distribute only"direct requests" and do not address a
load imbalance problem resulting from too many requests for hot
objects being simultaneously forwarded to the same proxy server.
The present invention addresses such a need.
[0005] Cooperative caching, or remote caching, has been used in
distributed file systems to improve system performance
(see"Cooperative caching: Using Remote Client Memory to Improve
File System Performance," by M. D. Dahlin et al., Proc. of 1st
Symp. on Operating Systems Design and Implementation, pp. 1-14,
1994). Here, the file caches of a collection of workstations
distributed on a LAN are coordinated to form a more effective
overall file cache. Each workstation caches not only objects
referenced by local requests but also objects that may be
referenced by requests from a remote workstation. Upon a local
cache miss, a local request can be sent to other client
workstations where a copy can be obtained, if found. Otherwise, the
object is obtained from the object server. The emphasis here is
mainly how to maintain cache coherency in the face of updates and
how to maintain cache hit ratios by moving a locally replaced
object to the cache memory of another workstation. There is no
dynamic load balancing.
[0006] Cooperative caching is also used in collective proxy cache
servers to reduce the access time. Upon a cache miss, instead of
going directly to the originating web server potentially through a
WAN, a cache server may forward the request to obtain the object
from a cooperating cache server in a LAN or a regional area
network. For example, upon a local cache miss in the SQUID system,
a cache server multicasts a request (using the Internet Cache
Protocol (ICP)) to a set of other cache servers ( see"Squid
Internet Object Cache", by D. Wessels et al.,
http://squid.nlanr.net/). If their caches contain the requested
object, these cooperating cache servers reply with a message
indicating such. The requested object is then obtained from the
cooperating cache server which responded first to the request,
instead of from the original web server on the Internet. However,
if none replies after a time-out period, then the requested object
will be fetched from the originating web server. Load imbalances
can occur at a cache server due to forwarded requests.
[0007] Instead of multicasting, the CRISP system uses a logical
central directory to locate an object cached on another proxy
server (see"Directory Structures for Scaleable Internet Caches", S.
Gadde et al., Technical Report CS-1997-18, Dept. of Computer
Science, Duke University, 1997). Here, upon a cache miss, a cache
server asks the directory server for the obj ect. With central
knowledge of the caches object storage, the directory server sends
such a request to the server whose cache includes the object. If
found, the object is then sent to the requesting server while the
original server continues to cache the object. If no cache has a
copy of the requested object, the requesting server obtains the
object from the originating web server through the Internet
(potentially through a WAN). Again, this can create a load
imbalance at the cache server due to subsequent requests forwarded
to this cache server.
[0008] Yet another way to locate an object on a cooperating cache
server is through a hash function. An example is the Cache Array
Routing Protocol (CARP) (see V. Valloppillil and K. W. Ross, "Cache
Array Routing Protocol v 1.0," Internet Draft,
http://ircache.nlanr.net/Cache/ICP/draft- -vinod-carp-v1-03.txt,
February 1998). In CARP, the entire object space is partitioned
among the cooperating cache servers, with one partition for each
cache server. When a request is received by a cache server from a
configured client browser, a hash function is applied to a key from
the request, such as the URL or the destination IP address, to
identify the partition. If the hash partition is the assigned to
requesting cache server, then the request is serviced locally.
Otherwise, it is forwarded to the proper cache server in the
identified partition.
[0009] SQUID, CRISP and CARP use the caches of other proxy servers
to reduce the possibility of having to go through the WAN for a
missed object. They differ in the mechanism for locating a
cooperating cache server whose cache may contain a copy of the
requested object. Each cache server services two kinds of requests:
direct requests and forwarded requests. Direct requests are those
made directly from the browsers connected to the proxy server.
Forwarded requests are those made by cooperating cache servers
whose caches do not have the requested objects. In any event,
depending on the types of objects a proxy server caches at a given
moment, its CPU could be overloaded because it is busy serving both
direct and forwarded requests.
SUMMARY OF THE INVENTION
[0010] In accordance with the aforementioned needs, the present
invention is directed to a method and system for balancing the load
across a collection of cache servers that process both direct and
forwarded requests by shifting some or all forwarded requests to a
less loaded cache server.
[0011] For example, in a system including a collection of
cooperating proxy cache servers, a request can be forwarded to
another cooperating server if the requested object cannot be found
locally. Instead of fetching the object from the originating web
server through the Internet, a cache server can obtain a copy from
a cooperating cache server in a local area network or an intranet.
The average response time for access to an object can be
significantly improved by the cooperating cache server. However,
due to reference skew, some objects can be in high demand by all
the clients. As a result, the proxy cache servers that contain
those hot objects can become overloaded by forwarded requests
coming from other proxy cache servers, creating a performance
bottleneck. According to the present invention, we propose a load
balancing method for a collection of cooperating proxy cache
servers by shifting some or all of the forwarded requests from an
overloaded cache server to a less loaded one.
[0012] An example of a cache server load balancing method in
accordance with the present invention includes the steps of:
receiving forwarded requests from a cooperating cache server in
response to a cache miss for an object on the cooperating cache
server; and shifting one or more of the forwarded requests for the
object between cooperating cache servers based on a load condition
and a forwarding frequency for the object.
[0013] The present invention also includes features for
periodically monitoring the load condition on and the forwarding
frequency to the owning cache server; and proactively shifting one
or more subsequent forwarded requests for the cached object from
the owning cache server to one or more of the cooperating cache
servers, in response to the monitoring. Alternatively, the shifting
step further includes the step of checking the load condition and
forwarding frequency, in response to the receipt of a forwarded
request. In one example, the load condition of the cooperating
cache server is a weighted sum of a count of said forwarded
requests, and a count of direct requests to said cooperating cache
server. In another example, the cache information is maintained at:
each object level; or a partition of objects level.
[0014] The present invention also includes various implementations
for performing the load balancing, including both centralized and
distributed environments and various hybrids thereof. For example,
a distributed load monitor can be used for monitoring and
maintaining a local load condition, the forwarding frequency and
ownership information for cached objects on each cooperating cache
server. The cooperating cache servers can periodically exchange and
maintain one or more of: the load condition information; the
forwarding frequency; and the ownership information. For example,
the cooperating cache servers can exchange information by
piggybacking one or more of: the load condition information; the
forwarding frequency; and the ownership information, with one or
more of the forwarded requests and responses.
[0015] In another example, an overloaded cooperating cache server
can identify a less loaded cooperating cache server; and
communicate a shift request and a copy of the cached object to the
less loaded cooperating cache server (which then caches the
object), so that subsequent requests for the object will not be
forwarded. Alternatively, an overloaded cooperating cache server
can communicate the shift request to the less loaded cooperating
cache server, which then obtains a copy of the object from an
originating object server, in response to the shift request. In yet
another alternative, the owning cache server can multicast the
shift request message to one or more of the other cooperating cache
servers so that subsequent forward requests will be shifted.
[0016] In a fully distributed implementation of the present
invention, the cooperating cache servers can each include a
distributed load monitor for monitoring and locally maintaining
load conditions, and also can maintain the forwarding frequency and
ownership information in a local copy of a caching table or by
means of a hashing function. The cooperating cache servers can
modify the ownership information by means of the local copy of the
caching table or the hash function.
[0017] The present invention includes still other features for
modifying the ownership for the object to a shared ownership
between at least two of the cooperating cache servers and
forwarding subsequent object requests to one or more less loaded
shared owners of the object. If a decrease in the load condition
for a shared object is detected, the shared ownership can be
merged, in response to the decrease in the load condition.
[0018] In yet another example, the shifting of one or more of the
forwarded requests based on the load conditon an the forwarding
frequency can be accomplished by communicating a copy of the object
from the owning cache server to one or more of the cooperating
cache servers, so that subsequent requests will not be forwarded
(as long as the object remains in the recipient's cache).
[0019] An example of a centralized environment in accordance with
the present invention includes: a centralized logical load monitor
for maintaining the forwarding frequency and the load condition for
the cooperating cache servers. The load monitor can include a
logical directory server for maintaining a load table for
monitoring the load on the cache servers and a caching table (or
hash function) for monitoring the forwarding frequency and locating
objects. The directory server receives requests for object
locations in other cache servers for a locally missed object and
forwards requests for locally missed objects. The directory server
load balances requests among the cooperating cache servers by
manipulating the caching table based on the load and the forwarding
frequency for a given object, in response to the requests for
object locations.
BRIEF DESCRIPTION OF THE DRAWINGS
[0020] These and other features, aspects, and advantages of the
present invention will become better understood with reference to
the following description, appended claims, and accompanying
drawings wherein:
[0021] FIG. 1a shows an example of a system in a block diagram form
employing a collection of proxy cache servers, wherein a
centralized load balancing logic according to the present invention
can be applied;
[0022] FIG. 1b shows another example of a system in a block diagram
form employing a collection of proxy cache servers, where a
distributed load balancing logic according to the present invention
can be applied;
[0023] FIGS. 2a-b show examples of data formats for two tables that
can be maintained by the load monitor depicted in FIGS. 1a-b;
[0024] FIG. 3 shows an example of a logic flow for the load monitor
in response to a request from a cache server because of a cache
miss; and
[0025] FIG. 4 shows an example of a logic flow for a cache server
in response to a request for an object.
DETAILED DESCRIPTION
[0026] Examples of the load balancing logic of the present
invention will be described for both centralized and distributed
architectures. FIG. 1a shows an example of a block diagram of a
system employing a collection of proxy cache servers, where a
centralized load balancing logic proposed in this invention can be
applied. As depicted, the system includes a collection of proxy
cache servers 150. Although only a single level of cache server is
depicted, there could be a hierarchy of cache servers 150. As is
conventional, these proxy cache servers are connected with each
other through a local area network (LAN) or a regional area network
or intranet 140. Each cache server 150 is also connected to a wide
area network (WAN) or the Internet 110. Through the WAN, these
proxy cache servers can reach 115 the originating web servers for
objects that cannot be found locally on their own caches.
[0027] According to the present invention a logical load monitor
120 includes a load balancing logic 130 for monitoring the load
conditions and forwarding frequency (FIG. 2a) of the cooperating
cache servers 150 and provides load balancing for them. As will be
described below, various load monitor 120 features can: reside in
one or more of the cache servers; be duplicated and distributed
among the cache servers; or reside in another dedicated system such
as a personal computer (PC) server or workstation. In a centralized
system configuration, the load monitor 120 can perform a central
directory function in directing forwarded requests 125 to the cache
servers. One or more browsers 160 can be configured to connect to
each cache server 150. Direct requests 155 are sent from the
clients such as computers running conventional browsers 160 to the
configured cache server 150. If the requested object can be found
locally, then it is returned to the browser. Otherwise, the cache
server 150 communicates a message to the load monitor 120. Various
example implementations of the load monitor 120 will be described
in more detail below. If no load imbalance condition or trend
exits, the load monitor 120 then forwards the request 125 to the
cache server 150 that owns the requested object. The owning cache
server then sends the requested object to the requesting cache
server, e.g., via the LAN 140.
[0028] If an actual load imbalance is identified, or predicted
based on a loading trend, the load monitor 120 initiates a shifting
of forwarded requests from the overloaded cache server to one or
more underloaded (or less loaded) servers. As will be described in
more detail below, the shifting of ownership can be based on the
load condition of the servers 150 and the forwarding frequency, as
well as other factors.
[0029] FIGS. 2a-b shows examples of data formats of two tables
maintained by the load monitor. As depicted, the tables include a
load table 102, and a caching table 101. One skilled in the art
will appreciate that a single table, or various other data
structures could alternatively or equivalently be used. The load
table 102 includes the load condition 1021 of each (A,B,C . . .
1022) cache server 150 so that overloaded and underloaded servers
can be identified. As is conventional, load conditions 1021 can be
updated periodically by probing each cache server. The load of a
cache server can be a weighted sum of the number of forwarded
requests and the number of direct requests. An overloaded cache
server 150 can be identified by any conventional techniques, e.g.,
the load monitor can compute the mean load of all proxy cache
servers in past intervals. Overloaded cache servers can be those
with loads exceeding a threshold above the mean load. According to
the present invention, load balancing takes into account the amount
of overloading as well as the load due to the forwarding frequency
1011 of the cached objects. This way, the load monitor can decide
whether or not to continue shifting some or all forwarded requests
from an overloaded cache server C 10213 to an underloaded server A
10211. The caching table 1010 includes the forwarding frequency
1011 and ownership 1012 information of an object or a partition of
objects. As will be discussed below, the ownership can be single as
in A 10122, or shared 10121, 10123 among two or more cooperating
cache servers. The forwarding frequency 1011 represents the number
of times a request for an object has been forwarded through the
load monitor. In addition to the forwarding frequency 1011, the
caching table 101 can also maintain a timestamp 1013, indicating
the most recent time a request for an object was forwarded.
Further, the caching information for an object or a partition 1010
can include a forwarding frequency over a given time period
(count/time) for the object ID or partition ID 1010 through the
load monitor 120. Object partitions 1010 can alternatively be based
on a hash function on object identifiers, or can be based on the
directory structures that objects are organized by on the web
servers. In the case of a partition, any object belonging to a
partition will be forwarded by the load monitor. The shifting of
ownership can be based on the load condition of the servers, the
forwarding frequency 1011 and other information such as the time
stamp information.
[0030] FIG. 3 shows an example of a logic flow for steps taken by
the load monitor 120 in response to a request 125 from a cache
server 150 because of a cache miss. As depicted, in step 201, it
checks to see if the requested object/partition can be found in the
caching table. If not, in step 202, a new entry is created for the
object/partition and a cache server is assigned as its owner. After
the entry is located in the caching table, in step 203, the
forwarding frequency 1011 is updated, e.g., incremented by 1. The
load monitor then examines the load table 102 to see if the owner
is currently overloaded (and that the forwarding frequency 1011 is
a significant contributor thereto), in step 204. If yes, in step
205, the load monitor finds an underloaded (or less loaded) cache
server and assign it as the new 10122 (or shared) owner 10122 of
the requested object. The ownership information 1012 for the object
in the caching table 101 is updated accordingly. Those skilled in
the art will appreciate that the logic flow could comprise a shared
10123 or hierarchical ownership 1012 in the caching table 101 or
other data structure employed. The request (possibly with a copy of
the requested object) can then be forwarded 125 to a new sole 10122
(or shared 10123) owner, in step 206. Alternatively, the new owner
can be requested to obtain 115 an object copy from the originating
object server, e.g., via the Internet 110. Those skilled in the art
will appreciate that the load checking step 204 can be performed
proactively, i.e., periodically or in response to an identified
overload or overload trend 1021--due at least in part to a high
forwarding frequency 1011--for a given object id/partition id 1010
and cache server (ownership 1012). If so, then in step 205, the
load monitor finds an underloaded (or less loaded) cache server,
assigns it as the new (or shared) owner of the requested object,
and possibly sends a copy of the object to the new (or shared)
owner as above. Conversely, if a shared ownership model is used, in
step 208, when the load condition 10211 and forwarding frequency
10111 for a shared ownership object (p 10101) drops below a
predetermined threshold, in step 209, the shared ownership (B, A
10121) can be merged to a single ownership and one of the copies
purged from one of the cache servers A 10121, e.g., to make room
for another hot object.
[0031] FIG. 4 shows an example of a logic flow for a cache server
when a request for an object is received, either directly 155 from
a browser 160 or forwarded 125 from the load monitor 120. As
depicted, in step 301, it first checks to see if the requested
object can be found locally in its cache. If yes, in step 302, it
returns the object and the process ends, in step 306. Otherwise, in
step 303, it checks to see if the request is a direct request or a
forwarded request. If it is a direct request, in step 304, the
request is sent to the load monitor and the process ends, in step
306. On the other hand, if the request is a forwarded request, in
step 305, the cache server will fetch the object from the
originating web server and return the object. The process then
ends, in step 306.
[0032] Referring now to FIGs. 1a and 2a-b, assume for example, a
browser 160 connecting to a cache server C 10223 requests 155 an
object p 10101. From the caching table 101, it can be seen that
object p 10101 is not cached on server C, but it is cached on
("owned" by) cache server B (assuming B, A 10121 is initially only
solely designated by B). In response to a cache miss on object p,
server C 10223 sends a request to the load monitor 120 for object
p. Depending on the load condition 10212 and forwarding frequency
1011 of requests for p 10101 on server B, the load monitor may
forward the request to server B, asking it to send a copy of object
p to server C. Or, if server B is currently overloaded or is
trending as such, the load monitor might shift the forwarded
request by finding an underloaded (or less loaded) server to serve
as a new (or shared as in B, A 10121) owner of object p. The
request is then forwarded to the new (or shared e.g., A) owning
server for the object. Note that even after the transfer of
ownership, a copy of object p is still on server B's cache and can
still serve direct requests coming to server B. However, in this
example, all future forwarded requests for object p (or perhaps
some, in the case of a shared ownership) will be shifted to server
A. Alternatively, in the case of shared ownership B, A 10121,
future forwarded requests for object p 10101 can be sent to the
less loaded server.
[0033] Now that a load balancing method according to the present
invention has been described for a collection of proxy cache
servers where a logical central directory is used for locating an
object, various alternatives will be considered. The present
invention can be adapted to achieve load balancing for these
systems as well.
[0034] For example, the present invention can be configured to
perform load balancing for a collection of cooperating proxy cache
servers where each cache server 150 multicasts to a list of
cooperating cache servers to locate a copy of a locally missed
object. In this case, no specific ownership information need be
maintained anywhere in the system. However, there is also no
guarantee of finding an object from the cooperating cache servers,
either. Assume that a logical load monitor 120 is used to maintain
the load conditions 1021 of all proxy cache servers and share this
information with each cache server 150. The load balancing can be
achieved by excluding overloaded servers from the list of
cooperating servers to which a cache server multicasts its request
(also called a shift request). As a result, only less loaded cache
servers will receive forwarded requests 125.
[0035] Another alternative is a load balancing method for a
collection of cooperating proxy cache servers where a hash function
is used to locate a copy of a locally missed object. In this case,
the object space can be partitioned among the cooperating proxy
cache servers 150, with one partition for each cache server. In
order to achieve load balancing by shifting forwarded requests, one
can change the hash function so that forwarded requests will not go
to overloaded servers. One preferred approach is to hash the object
space into a large number of buckets, much larger than the total
number of proxy cache servers. These hash buckets are then assigned
to the cache servers, with the goal of balancing the loads among
them. Periodically, one can move one or more hash buckets from one
overloaded server to an underloaded server, effectively changing
the hash function.
[0036] In either case, the load condition of the cooperating cache
server can factor in the forwarding frequency directly into the
calculated load condition. For example, the load condition can be a
weighted sum of a count of said forwarded requests, and a count of
direct requests to said cooperating cache server. Alternatively,
the load monitor could separately maintain the overal forwarding
frequency for each cooperating cache server.
[0037] Referring now to FIGS. 1b and 2a-b, yet another alternative
is a load monitor 120 that is distributed, i.e., wherein some or
all the load monitor is duplicated across the cache servers 150. In
one example, the distributed load monitor includes local load
condition information 1021 (and as described below, possibly the
load conditions of all (A,B,C, . . . 1022)) of the cooperating
cache servers 150. The distributed load monitor 120' preferably
also includes the caching table 101 with the forwarding frequency
1011 and ownership 1012 information for each object id/partition id
1010. Alternatively, a hashing function, for example as described
above, could be distributed and stored in the cache servers. Load
condition information 1021 and/or caching information 101: can be
exchanged periodically; when there is a change in status (ownership
or significant change in load condition); or piggybacked with cache
forwarding requests and responses. Load condition 1021 information
could also have a time stamp (not shown) associated with it for
tracking or other purposes.
[0038] Here, if a cache server 150 has a cache miss, the local load
monitor 120' looks up the ownership of the requested object in its
local caching table 101 and forwards the request, to the owning
cache server. Alternatively, the hash function could be applied to
a key from the request, such as the URL or the destination IP
address, to identify the partition and the request then forwarded
to the correct cache server. When the forwarded request (i.e., from
a cache server who had a cache miss) is received, the owning cache
server identifies it as a forwarded request (e.g., by identifying
it as from another cache server as opposed to a client) and updates
its forwarding frequency 1011 information as applicable (FIG. 3,
step 203). If an overload trend or condition is indicated (step
204), the owning cache server can respond to the requesting cache
server with a shift request and a copy of the cached object.
Alternatively, the requesting cache server can obtain a copy from
the originating object server via an intranet, WAN or Internet 110.
In either case, when the forwarding server caches a copy of the
object, this server will no longer issue forward requests (steps
301, 302) as long as it remains in the cache, thus proportionally
reducing the load on the owning server. In addition, the owning
cache server can multicast a shift request message to one or more
of the other cooperating cache servers 150 so that subsequent
forward requests will be shifted, e.g., by updating their local
copy of the caching table or modifying the hash function (step
205). At this point, other cache servers can forward their requests
to the new owner (or to the least loaded owner of two or more cache
servers 150 if ownership is shared) as indicated in their local
copy of the caching table 101. When the original cache owner's load
has decreased to an acceptable level (step 204), e.g., as indicated
by a threshold, the shared ownership information can be merged to
its original state (e.g., B,A 10121-->B).
[0039] In the case that the load condition information 1021 for all
cache servers ( A,B,C . . . 1022) is fully distributed, the
requesting cache server could proactively check the load condition
(and associated time stamp) of the owning server (step 204), i.e.,
before forwarding the request. If overloaded, the requesting server
could request a copy of the object from the owning server (or from
the originating server via the intranet or Internet 110) and
possibly a load condition confirmation. The owning cache server
could update its caching table 101 or modify the hash function to
indicate the new shared ownership (step 205). The requesting server
(or the owning server) could then multicast a message to all other
cache servers 150 indicating the new shared ownership of the object
and possibly include an updated load condition. At this point,
other cache servers would update their caching tables 101 or modify
the hash function to indicate the new shared ownership (step 202),
and can forward their requests (step 206) to the least loaded owner
of two or more cache servers 150 sharing ownership as indicated in
their local copy of the caching table 101. When a shared cache
owner's load has decreased to an acceptable level (steps 204 and
208), e.g., as indicated by a threshold, the ownership information
can be merged to its original state, in step 209.
[0040] A preferred embodiment of the present invention includes
features that can be implemented as software tangibly embodied on a
computer program product or program storage device for execution on
a processor (not shown) provided with cache server 150 or other
computer embodying the load monitor 120, such as in the centralized
model described. For example, software implemented in a popular
object-oriented computer executable code such as JAVA provides
portability across different platforms. Those skilled in the art
will appreciate that many other compiled or interpreted,
procedure-oriented and/or object-oriented (OO) programming
environments, including but not limited to REXX, C, C.sup.++ and
Smalltalk can also be employed.
[0041] Those skilled in the art will also appreciate that methods
of the present invention may be the software may be embodied on a
magnetic, electrical, optical, or other persistent program and/or
data storage device, including but not limited to: magnetic disks,
Direct Access Storage Devices (DASD), bubble memory; tape; optical
disk formats such as CD-ROMs and DVD; and other persistent (also
called nonvolatile) storage devices such as core, ROM, PROM, flash
memory, or battery backed RAM. Those skilled in the art will
appreciate that within the spirit and scope of the present
invention, one or more of the components instantiated in the memory
of the server 120' could be accessed and maintained directly via
disk (not shown), the network, another server, or could be
distributed across a plurality of servers.
[0042] While we have described our preferred embodiments of our
invention with alternatives, it will be understood that those
skilled in the art, both now and in the future, may make various
improvements and enhancements which fall within the scope of the
claims which follow. These claims should be construed to maintain
the proper protection for the invention first disclosed.
* * * * *
References