Right, that’s it. I think what happened is the following: all the nodes generated some garbage that put them very close to the threshold for a full GC in the first few runs of the program (when you cached the RDDs), but on the subsequent queries, only a few nodes are hitting full GC per query, so every query sees a slowdown but the problem persists for a whille. You can try manually forcing a GC on the nodes like this after you do your loading:
sc.parallelize(1 to numNodes, numNodes).foreach(x => System.gc()) Where numNodes is your number of nodes. (Actually it’s also okay to just make this higher, System.gc() returns fast when there’s no GC to run.) Matei On Mar 11, 2014, at 7:12 AM, Koert Kuipers <ko...@tresata.com> wrote: > hey matei, > most tasks have GC times of 200ms or less, and then a few tasks take many > seconds. example GC activity for a slow one: > > [GC [PSYoungGen: 1051814K->262624K(1398144K)] 3789259K->3524429K(5592448K), > 0.0986800 secs] [Times: user=1.53 sys=0.01, real=0.10 secs] > [GC [PSYoungGen: 786935K->524512K(1398144K)] 4048741K->4048762K(5592448K), > 0.1132490 secs] [Times: user=1.70 sys=0.01, real=0.11 secs] > [Full GC [PSYoungGen: 524512K->0K(1398144K)] [PSOldGen: > 3524250K->2207344K(4194304K)] 4048762K->2207344K(5592448K) [PSPermGen: > 56545K->54639K(83968K)], 7.7059350 secs] [Times:\ > user=7.71 sys=0.00, real=7.70 secs] > > > so looks like i am hit by stop-the-world gc? > > > On Mon, Mar 10, 2014 at 7:00 PM, Koert Kuipers <ko...@tresata.com> wrote: > hey matei, > it happens repeatedly. > > we are currently runnning on java 6 with spark 0.9. > > i will add -XX:+PrintGCDetails and collect details, and also look into java 7 > G1. thanks > > > > > > > On Mon, Mar 10, 2014 at 6:27 PM, Matei Zaharia <matei.zaha...@gmail.com> > wrote: > Does this happen repeatedly if you keep running the computation, or just the > first time? It may take time to move these Java objects to the old generation > the first time you run queries, which could lead to a GC pause that also > slows down the small queries. > > If you can run with -XX:+PrintGCDetails in your Java options, it would also > be good to see what percent of each GC generation is used. > > The concurrent mark-and-sweep GC -XX:+UseConcMarkSweepGC or the G1 GC in Java > 7 (-XX:+UseG1GC) might also avoid these pauses by GCing concurrently with > your application threads. > > Matei > > On Mar 10, 2014, at 3:18 PM, Koert Kuipers <ko...@tresata.com> wrote: > >> hello all, >> i am observing a strange result. i have a computation that i run on a cached >> RDD in spark-standalone. it typically takes about 4 seconds. >> >> but when other RDDs that are not relevant to the computation at hand are >> cached in memory (in same spark context), the computation takes 40 seconds >> or more. >> >> the problem seems to be GC time, which goes from milliseconds to tens of >> seconds. >> >> note that my issue is not that memory is full. i have cached about 14G in >> RDDs with 66G available across workers for the application. also my >> computation did not push any cached RDD out of memory. >> >> any ideas? > > >
