Changes between Version 1 and Version 2 of LowLatency
- Timestamp:
- Apr 25, 2007, 1:21:16 PM (18 years ago)
Legend:
- Unmodified
- Added
- Removed
- Modified
-
LowLatency
v1 v2 1 1 = Low-latency computing = 2 2 3 3 BOINC was originally designed for high-throughput computing, and one of its basic design goals was to minimize the number of scheduler RPCs (in order to reduce server load and increase scalability). In particular, when a client requests work from a server and there is none, the client uses exponential backoff, up to a maximum backoff off 1 day or so. This policy limits the number of scheduler requests to (roughly) one per job. However, this backoff policy is inappropriate for '''low-latency''' computing, by which we main projects whose tasks must be completed in a few minutes or hours. Such projects require a '''minimum connection rate''', rather than seeking to minimize the connection rate. 4 4 5 5 For example, if you need to get batches of 10,000 jobs completed with 5 minute latency, and each job takes 2 minutes of computing, you'll need to arrange to get 10,000 scheduler requests every 3 minutes (and you'll need a server capable to handling this request rate). 6 6 7 7 8 == = The minimum connection rate ===9 8 == The minimum connection rate == 9 Suppose that, at a given time, the project has N hosts online, and that each host has 1 CPU that computes at X FLOPS. 10 10 11 11 Suppose that the project's work consists of 'batches' of M jobs. Each batch is generated at a particular time, and all the jobs must be completed within time T. For simplicity, assume that a batch is not created until the previous batch has been completed, and that each has is given at most one job from each batch. Suppose that each job takes Y seconds to complete on the representative X-FLOPS CPU. … … 18 18 19 19 20 === How to do low-latency computing === 21 The key component in the above is the ability to control Z, the time between requests for a given host. Starting with version 5.6 of the BOINC client, it is now possible to control this: each scheduler reply can include a tag 22 20 == How to do low-latency computing == 21 The key component in the above is the ability to control Z, the time between requests for a given host. Starting with version 5.6 of the BOINC client, it is now possible to control this: each scheduler reply can include a tag 23 22 24 23 {{{<next_rpc_delay>x</next_rpc_delay>}}} 25 telling the client when to contact the scheduler again. By varying this value, a project can achieve a rate of connection requests necessary to achieve its latency bounds. The current BOINC scheduler code has no support for sending this tag, or for figuring out what its value should be. If you want to do low-latency computing, the scheduler must be modified and extended as follows:26 24 25 telling the client when to contact the scheduler again. By varying this value, a project can achieve a rate of connection requests necessary to achieve its latency bounds. The current BOINC scheduler code has no support for sending this tag, or for figuring out what its value should be. If you want to do low-latency computing, the scheduler must be modified and extended as follows: 27 26 28 27 * Keep track of how many active hosts you have (this will change over time).
