Distributed traffic shaping across a cluster

• US 7,760,641 B2
• Filed: 07/10/2006
• Issued: 07/20/2010
• Est. Priority Date: 07/10/2006
• Status: Expired due to Fees

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Gilfix, Michael A.
• Primary Examiner(s): Shah; Chirag G
• Assistant Examiner(s): YOUNG, STEVE R
• Application Number:
  US11/456,462
• Publication Number:
  US 20080008095A1
• Time in Patent Office: 1,471 Days
• Field of Search: None
• US Class Current: 370/235
• CPC Class Codes: H04L 47/10 Flow control; Congestion control
• CPC Class Codes: H04L 47/19 at layers above the network layer network arrangements for networked applications for scheduling or organising the servicing of application requests H04L67/60
• CPC Class Codes: H04L 47/215 using token-bucket
• CPC Class Codes: H04L 67/1001 for accessing one among a plurality of replicated servers
• CPC Class Codes: H04L 67/1008 based on parameters of servers, e.g. available memory or workload monitoring of computer activity G06F11/30
• CPC Class Codes: H04L 67/1014 based on the content of a request
• CPC Class Codes: H04L 67/1023 based on a hash applied to IP addresses or costs
• CPC Class Codes: H04L 67/1034 Reaction to server failures by a load balancer
• CPC Class Codes: H04L 67/14 Session management for real-time applications in data packet communications networks H04L65/1066
• CPC Class Codes: H04L 67/63 Routing a service request depending on the request content or context

Abstract

A network is provided herein comprising a plurality of network resources, and at least one network cluster having a plurality of cluster members. Each member of the cluster may be configured for utilizing one or more of the network resources and for tracking usage thereof. For example, each member of the cluster may include one or more token buckets for tracking that member'"'"'s usage of the network resources. At least one member of the cluster (i.e., a “reservation coordinator”) may include a first set of computer-executable instructions for receiving network traffic destined for a particular network resource at a first rate (i.e., a maximum average sustained rate). In addition, the reservation coordinator may include a second set of computer-executable instructions for distributing the first rate among at least a subset of the cluster members. In some cases, each member of the cluster may include a third set of computer-executable instructions for changing how the first rate is distributed among members of the subset. For example, the third set of computer-executable instructions may include instructions for attempting to reserve a substantially larger or smaller portion of the first rate.

Claims

1. A method for controlling the distribution of network traffic sent to a network cluster having a plurality of cluster members, the method comprising:

maintaining, in a reservation coordinator cluster member of the network cluster, a hierarchical cluster rate limit tree data structure for the entire network cluster and a separate hierarchical member rate limit tree data structure for each cluster member in the plurality of cluster members, wherein the cluster rate limit tree data structure tracks rate limits and rate consumption across all of the plurality of cluster members, and wherein each separate member rate limit tree data structure tracks rate limits and rate consumption for individual applications and application operations of an associated cluster member;

receiving network traffic destined for a particular application at a first rate, the network traffic comprising a plurality of client requests for the application, and wherein each of the cluster members is an application server that concurrently provides the application; and

distributing the first rate among a subset of the cluster members based on the cluster rate limit tree data structure and the separate member rate limit tree data structures such that each cluster member in the subset has a portion of the first rate and a workload corresponding to the received network traffic is shared amongst the cluster members in the subset of the cluster members, wherein;

each cluster member within the subset maintains a token bucket for the particular application,

the portion of the first rate distributed to each cluster member within the subset corresponds to a regeneration rate for a respective token bucket, and

each of the member rate limit tree data structures comprise nodes in a hierarchical arrangement and rate limits associated with the nodes such that a rate limit of a parent node in the hierarchical arrangement equals a sum of rate limits associated with child nodes of the parent node.

2. The method as recited in claim 1, further comprising distributing burst among the subset of cluster members such that each cluster member in the subset is allocated a portion of the burst.

3. The method as recited in claim 2, wherein a sum of the portions of the first rate distributed to each cluster member within the subset equals the first rate, and wherein a sum of the portions of the burst distributed to each cluster member within the subset equals a total burst output for the network cluster.

4. The method as recited in claim 3, wherein a substantially equal amount of rate and burst is distributed to each cluster member within the subset.

5. The method as recited in claim 3, wherein a portion of the first rate and a portion of burst distributed to one cluster member within the subset differs from other portions of the first rate and other portions of burst distributed to other cluster members within the subset.

6. The method as recited in claim 3, wherein each client request in the plurality of client requests has an associated weight, and wherein the total burst output divided by a largest portion of burst distributed to any of the cluster members within the subset is greater than or equal to a largest weight assigned to the plurality of client requests.

7. The method as recited in claim 3, further comprising redistributing the total burst output among active members of the subset, if one or more cluster members of the subset fails or is taken off-line.

8. The method as recited in claim 1, further comprising:

directing one of the client requests to one of the cluster members in the subset; and

accessing a local token bucket stored within the one cluster member for tracking usage of the particular application, wherein the local token bucket contains a current number of tokens and is configured for accepting additional tokens at a second rate equal to the regeneration rate assigned to the local token bucket.

9. The method as recited in claim 8, further comprising consuming a number of tokens from the local token bucket, the number being equal to a weight assigned to the one client request, if the current number of tokens contained within the local token bucket is greater than or equal to the weight assigned to the one client request.

10. The method as recited in claim 9, further comprising attempting to increase the regeneration rate assigned to the local token bucket regardless of whether or not tokens were consumed from the local token bucket.

11. The method as recited in claim 10, wherein attempting to increase the regeneration rate comprises:

sending a reservation request to an estimation algorithm configured for determining when to satisfy a reservation request; and

receiving a reservation amount from the estimation algorithm, wherein the reservation amount is based on whether or not the one client request was received in a burst of network traffic.

12. The method as recited in claim 11, wherein if the one client request was not received in a burst of network traffic, receiving a reservation amount comprises:

receiving a zero reservation amount from the estimation algorithm, if the reservation request is sent within a designated silence period; and

receiving a non-zero reservation amount from the estimation algorithm, if the reservation request is sent after the designated silence period ends.

13. The method as recited in claim 12, wherein if the one client request was received in a burst of network traffic, receiving a reservation amount comprises:

determining a current rate consumption trend for all network traffic received by the one cluster member for the particular network service during a current time period;

receiving a zero reservation amount from the estimation algorithm, if;

(a) the reservation request is sent within a designated silence period, or (b) the current rate consumption trend for the one cluster member is decreasing or steady and the amount of rate distributed to the one cluster member is greater than zero; and

receiving a non-zero reservation amount from the estimation algorithm, if the reservation request is sent after the designated silence period ends.

14. The method as recited in claim 13, wherein upon receiving a non-zero reservation amount from the estimation algorithm, attempting to increase the regeneration rate further comprises:

sending the reservation request to a reservation algorithm configured for determining whether or not sufficient rate capacity remains within the cluster to satisfy the reservation request;

receiving a zero reservation amount from the reservation algorithm, if insufficient rate capacity remains within the cluster to satisfy the reservation request; and

receiving a non-zero reservation amount from the reservation algorithm, if sufficient rate capacity remains within the cluster to satisfy the reservation request.

15. The method as recited in claim 14, wherein upon receiving a non-zero reservation amount from the reservation algorithm, the method further comprises increasing the amount of rate distributed to the one cluster member by adding the non-zero reservation amount to the regeneration rate for the local token bucket.

22. The method of claim 8, wherein directing one of the client requests to one of the cluster members in the subset comprises:

traversing a member rate limiter tree data structure for the cluster member along a path from a cluster level to a child node of the member rate limiter tree data structure;

determining if each node along the path has sufficient capacity in the node'"'"'s corresponding token bucket to accept the client request;

directing the client request to the cluster member in response to a determination that each node along the path has sufficient capacity in the node'"'"'s corresponding token bucket to accept the client request; and

rejecting the client request in response to a determination that at least one node along the path does not have sufficient capacity in the node'"'"'s corresponding token bucket to accept the client request.

23. The method of claim 1, wherein a total of the rate limits of nodes in a member rate limiter tree data structure is equal to a portion of the first rate allocated to the cluster member in the cluster rate limiter tree data structure.

16. A network comprising:

a network cluster having a plurality of cluster members, each individually configured for concurrently providing a same application and for tracking usage thereof; and

a reservation coordinator cluster member for controlling a distribution of traffic rate limits amongst the plurality of cluster members, wherein;

the reservation coordinator cluster member maintains a hierarchical cluster rate limit tree data structure for the entire network cluster and a separate hierarchical member rate limit tree data structure for each cluster member in the plurality of cluster members,

the cluster rate limit tree data structure tracks rate limits and rate consumption across all of the plurality of cluster members,

each separate member rate limit tree data structure tracks rate limits and rate consumption for individual applications and application operations of an associated cluster member,

the network cluster receives network traffic destined for the application at a first rate, the network traffic comprising a plurality of client requests for the application,

the reservation coordinator distributes the first rate among a subset of the cluster members based on the cluster rate limit tree data structure and the separate member rate limit tree data structures such that each cluster member in the subset has a portion of the first rate and a workload corresponding to the received network traffic is shared amongst the cluster members in the subset of the cluster members,

each cluster member within the subset maintains a token bucket for the particular application,

the portion of the first rate distributed to each cluster member within the subset corresponds to a regeneration rate for a respective token bucket, and

each of the member rate limit tree data structures comprise nodes in a hierarchical arrangement and rate limits associated with the nodes such that a rate limit of a parent node in the hierarchical arrangement equals a sum of rate limits associated with child nodes of the parent node.

17. The network as recited in claim 16, wherein the second set of computer-executable instructions are further configured for distributing burst among active cluster members of the subset.

18. The network as recited in claim 16, wherein each cluster member of the cluster comprises a third set of computer-executable instructions for changing how the first rate is distributed among the cluster members of the subset.

19. The network as recited in claim 18, wherein the third set of computer-executable instructions comprise:

a fourth set of computer-executable instructions for attempting to reserve a larger portion of the first rate; and

a fifth set of computer-executable instructions for attempting to reserve a smaller portion of the first rate.

20. The network as recited in claim 19, wherein the fourth set of computer-executable instructions are executed by only one cluster member of the subset at any given time.

21. The network as recited in claim 19, wherein the fifth set of computer-executable instructions are executed by all cluster members of the subset at once.