Clustered dispersion of resource use in shared computing environments

• US 9,210,048 B1
• Filed: 03/31/2011
• Issued: 12/08/2015
• Est. Priority Date: 03/31/2011
• Status: Active Grant

• Current Assignee: Amazon Technologies, Inc. (Amazon.com, Inc.)
• Original Assignee: Amazon Technologies, Inc. (Amazon.com, Inc.)
• Inventors: Marr, Michael David, Lamoreaux, Tyson J.
• Primary Examiner(s): Chacko, Joe
• Application Number:
  US13/077,260
• Time in Patent Office: 1,713 Days
• Field of Search: 709/226
• US Class Current: 1/1
• CPC Class Codes: H04L 41/145 involving simulating, designing, planning or modelling of a network
• CPC Class Codes: H04L 43/0882 Utilisation of link capacity
• CPC Class Codes: H04L 47/70 Admission control; Resource allocation
• CPC Class Codes: H04L 47/827 Aggregation of resource allocation or reservation requests
• CPC Class Codes: H04L 49/20 Support for services
• CPC Class Codes: H04L 67/10 in which an application is distributed across nodes in the network software deployment G06F8/60; multiprogramming arrangements G06F9/46

Abstract

Host machines and other devices performing synchronized operations can be dispersed across multiple racks in a data center to provide additional buffer capacity and to reduce the likelihood of congestion. The level of dispersion can depend on factors such as the level of oversubscription, as it can be undesirable in a highly connected network to push excessive host traffic into the aggregation fabric. As oversubscription levels increase, the amount of dispersion can be reduced and two or more host machines can be clustered on a given rack, or otherwise connected through the same edge switch. By clustering a portion of the machines, some of the host traffic can be redirected by the respective edge switch without entering the aggregation fabric. When provisioning hosts for a customer, application, or synchronized operation, for example, the levels of clustering and dispersion can be balanced to minimize the likelihood for congestion throughout the network.

Claims

1. A computer-implemented method of managing traffic in a networked environment, comprising:

under control of one or more computer systems configured with executable instructions,

receiving a request to a provisioning component for the networked environment, the networked environment including a plurality of host servers connected by an aggregation fabric including layers of network switches, a path across the aggregation fabric to each host server involving a number of connections across the network switches;

determining a number of the host servers to be used in performing subsequent operations associated with a source of the request;

determining a level of subscription for at least a lowest layer of the aggregation fabric;

determining, for at least the lowest layer of the aggregation fabric, a number of network switches over which to disperse the determined number of host servers, the number of network switches being based at least in part on (i) the level of subscription for at least the lowest layer of the aggregation fabric and (ii) a topology of links associated with the aggregation fabric, the number of network switches at the lowest layer being at most equal to the determined number of host servers; and

in response to receiving one of the subsequent operations for processing, assigning a subset of the determined number of host servers to process the received subsequent operation,

wherein each of the determined host servers in the subset is operable to concurrently perform at least a portion of the received subsequent operation.

2. The computer-implemented method of claim 1, wherein the aggregation fabric is in one of an oversubscribed or non-oversubscribed state.

3. The computer-implemented method of claim 1, wherein the number of network switches over which to disperse the determined number of host servers is determined based at least in part upon a topology and subscription rate of links within the aggregation fabric.

4. The computer-implemented method of claim 1, wherein the number of network switches over which to disperse the determined number of host servers is determined based at least in part upon at least one of historical usage, a current usage, duty cycle, a time of day, or a day of a week.

5. The computer-implemented method of claim 1, wherein a plurality of host servers are operable to be connected to a single network switch or application-specific integrated circuit (ASIC) depending at least in part upon a subscription rate of the aggregation fabric.

6. The computer-implemented method of claim 5, wherein each network switch connected to a set of host servers is a top of rack (TOR) switch connected to a plurality of host devices in a server rack.

7. The computer-implemented method of claim 6, wherein multiple operations are capable of being assigned to the plurality of host servers in each server rack.

8. The computer-implemented method of claim 1, wherein a number of assigned host servers for an operation at a first network switch in the lowest layer is different than a number of assigned host servers for the operation at a second network switch in the lowest layer.

9. The computer-implemented method of claim 1, wherein the aggregation fabric connecting the plurality of host servers to an external network is a high radix interconnection network.

10. The computer-implemented method of claim 1, wherein at least one cluster of network switches for performing an operation is a network with re-arrangeably non-blocking bandwidth.

11. The computer-implemented method of claim 10, wherein at least one cluster of network switches for performing an operation is a Clos topology network.

12. The computer-implemented method of claim 1, wherein at least one cluster of network switches utilizes a network topology selected from a hierarchal router pair or an irregular ad hoc network topology.

13. A computer-implemented method of dispersing host machines for a customer in a data center, comprising:

under control of one or more computer systems configured with executable instructions,

receiving a request to provision a specified number of host servers for a customer, the host servers each being connected to one of a plurality of host server-connected switches of an aggregation network fabric comprising a plurality of network switches;

determining a level of subscription for at least a lowest layer of the aggregation network fabric;

determining a number of host server-connected switches over which to disperse the specified number of host servers, the number of host server-connected switches being based at least in part on the level of subscription and network topology for at least a lowest layer of the aggregation fabric; and

in response to receiving an operation for the customer, provisioning a subset of the specified number of host servers across the determined number of host server-connected switches to process the received operation for the customer,

wherein the received operation for the customer can be performed in parallel using the provisioned subset of host servers in the data center.

14. The computer-implemented method of claim 13, wherein the host servers communicate using conventional network communication protocols.

15. The computer-implemented method of claim 13, wherein the data center is connected using a high redix network.

16. The computer-implemented method of claim 13, wherein the network is a network with re-arrangeably non-blocking bandwidth.

17. The computer-implemented method of claim 16,

wherein the network is a Clos topology network.

18. The computer-implemented method of claim 13, wherein the network is a hierarchical router pair network.

19. The computer-implemented method of claim 13, wherein each host server-connected switch is a commodity top of rack (TOR) switch.

20. The computer-implemented method of claim 13, wherein the host servers communicate during a computation process using at least one application programming interface (API).

21. The computer-implemented method of claim 20, wherein the at least one API is implemented in at least one of a message passing interface (MPI) or an open message passing (OpenMP) interface.

22. The computer-implemented method of claim 13, wherein the host servers are further distributed across a plurality of sub-networks.

23. A system for managing data in a networked environment, comprising:

a hardware processor; and

memory including instructions that, when executed by the processor, cause the system to;

receive a request to a provisioning component for the networked environment, the networked environment including a plurality of host servers connected by an aggregation fabric including layers of network switches, a path across the aggregation fabric to each host server involving a common number of connections across the network switches;

determine a number of the host servers to be used in performing subsequent operations associated with a source of the request;

determine a level of subscription for at least a lowest layer of the aggregation fabric;

determine, for at least a lowest layer of the aggregation fabric, a number of network switches over which to disperse the determined number of host servers, the number of network switches being based at least in part on (i) the level of subscription for at least the lowest layer of the aggregation fabric and (ii) a topology of links associated with the aggregation fabric, the number of network switches at the lowest layer being at most equal to the determined number of host servers; and

in response to receiving one of the subsequent operations for processing, assign a subset of the determined number of host servers to process the received subsequent operation, wherein each of the determined host servers in the subset is operable to concurrently perform at least a portion of the received subsequent operation.

24. The system of claim 23, wherein the number of network switches over which to disperse the determined number of host servers is determined based at least in part upon a topology and subscription rate of links within the aggregation fabric.

25. The system of claim 23, wherein the number of network switches over which to disperse the determined number of host servers is determined based at least in part upon at least on of historical usage, a current usage, duty cycle, a time of day, or a day of the week.