http://dev.clojure.org/jira This file is an XML representation of an issue en-us 4.4 649 25-07-2011 http://dev.clojure.org/jira/browse/CLJ-862 Clojure <p>The code of pmap creates futures for the set of map operations it needs to perform with (map #(future %) coll), then acts on that sequence in a fashion obviously intended to keep only #CPUs+2 unfinished futures realized at any given time. It works exactly this way for non-chunked input seqs, but when pmap is passed a chunked seq, the seq of futures also becomes chunked. This causes an arbitrary number of futures to be realized as the #CPUs+2 window and chunk-size window interact.</p> <p>The doc-string for pmap doesn't promise any particular level of parallelism, but I think the inconsistent parallelism for chunked vs non-chunked input comprises a bug.</p> CLJ-862

pmap level of parallelism inconsistent for chunked vs non-chunked input

Defect Minor Open Unresolved Jim Blomo Marshall T. Vandegrift Sun, 23 Oct 2011 09:50:38 -0500 Sat, 20 Oct 2012 00:45:48 -0500 Release 1.1 Release 1.2 Release 1.3 Release 1.4 Reviewed Backlog 0 2 <p>Next person to take a deep look at pmap should probably also think through fork/join.</p> <p>fork/join is a Java 7 feature. Would a proposed patch need to be able to fall back to Java 5 features?</p> <p>Clojure/core folks can say more authoritatively, but I believe with the recent reduce enhancements that rely on jsr166 code, Clojure 1.5 will most likely require Java 6 or later, and Java 5 will no longer be supported.</p> <p>Spinning up more threads than CPU cores is not a good idea when the work is CPU bound. Currently (1.4) pmap uses an unbounded thread pool, so chunked sequences will create more threads than intended. The least invasive change is to use a fixed sized thread pool (ForkJoinPool being one example). pmap is differentiated from core.reducers by the fact that it is lazy. This implies a one-at-a-time ThreadPool.submit model, instead of the recursive fork/join model. Tradeoffs include:</p> <p>Enforce look-ahead even on chunked sequences:</p> <ul class="alternate" type="square"> <li>+ no threadPool changes</li> <li>- working against chunking, which is presumably being used for a reason</li> </ul> <p>Move to a fixed size thread pool:</p> <ul class="alternate" type="square"> <li>+ reduce contention for cpu-bound functions on chunked sequences</li> <li>- increase total realization time for io-bound functions</li> </ul> <p>Use ForkJoinPool for fixed thread pool (instead of newFixedThreadPool):</p> <ul class="alternate" type="square"> <li>+ automatic and dynamic parallelism</li> <li>- more complex setup (picking Java 6 vs 7 implementation, sharing pool with core.reducers)</li> </ul> <p>I think using a traditional fixed size thread pool is the right option. Most of the time all of pmap's results will be realized, so I don't think it's worth saving work by being strict about the look-ahead size. This is also the decision map has made. Since we're not using ForkJoin's main strength (recursive work queuing), I don't think it is worth setting it up in clojure.core.</p> <p>I'll use Agent/pooledExecutor as the fixed size thread.</p> <p>Let me know if I forgot or misunderstood anything.</p> <p>2012-05-28 pmap-chunking-862.diff uses a fixed size thread pool for pmap functions.</p> Global Rank Patch Code and Test