Thanks. So may I know what is your configuration for more/smaller executors on r3.8xlarge, how big of the memory that you eventually decide to give one executor without impact performance (for example: 64g? ).
From: Sven Krasser [mailto:kras...@gmail.com] Sent: Friday, April 24, 2015 1:59 PM To: Dean Wampler Cc: Shuai Zheng; user@spark.apache.org Subject: Re: Slower performance when bigger memory? FWIW, I ran into a similar issue on r3.8xlarge nodes and opted for more/smaller executors. Another observation was that one large executor results in less overall read throughput from S3 (using Amazon's EMRFS implementation) in case that matters to your application. -Sven On Thu, Apr 23, 2015 at 10:18 AM, Dean Wampler <deanwamp...@gmail.com> wrote: JVM's often have significant GC overhead with heaps bigger than 64GB. You might try your experiments with configurations below this threshold. dean Dean Wampler, Ph.D. Author: Programming Scala, 2nd Edition <http://shop.oreilly.com/product/0636920033073.do> (O'Reilly) Typesafe <http://typesafe.com> @deanwampler <http://twitter.com/deanwampler> http://polyglotprogramming.com On Thu, Apr 23, 2015 at 12:14 PM, Shuai Zheng <szheng.c...@gmail.com> wrote: Hi All, I am running some benchmark on r3*8xlarge instance. I have a cluster with one master (no executor on it) and one slave (r3*8xlarge). My job has 1000 tasks in stage 0. R3*8xlarge has 244G memory and 32 cores. If I create 4 executors, each has 8 core+50G memory, each task will take around 320s-380s. And if I only use one big executor with 32 cores and 200G memory, each task will take 760s-900s. And I check the log, looks like the minor GC takes much longer when using 200G memory: 285.242: [GC [PSYoungGen: 29027310K->8646087K(31119872K)] 38810417K->19703013K(135977472K), 11.2509770 secs] [Times: user=38.95 sys=120.65, real=11.25 secs] And when it uses 50G memory, the minor GC takes only less than 1s. I try to see what is the best way to configure the Spark. For some special reason, I tempt to use a bigger memory on single executor if no significant penalty on performance. But now looks like it is? Anyone has any idea? Regards, Shuai -- www.skrasser.com <http://www.skrasser.com/?utm_source=sig>