On Wed, Oct 17, 2012 at 2:59 PM, Dawid Weiss <dawid.we...@cs.put.poznan.pl> wrote: > Just a though -- we have a benchmark regression test that attempts to > estimate the 'minimum required memory before it hits an OOM' for a > known task (a fixed data set for clustering). Essentially we do a > binary search between not hitting an OOM and hitting an OOM. What this > brings us is that we know if memory requirements changed over time (or > if the JVM/ GC have changed over time...). > > This is a computationally expensive test (re-spinning the JVM, GC goes > insane on near-OOM conditions) but it's pretty useful. > > Again, this is just a thought, perhaps it'll be useful to investigate the > issue.
This would be really cool :) Then it's no longer this binary thing (does it fit in 512 MB or not), and we can see over time if we are taking more and more memory. Not sure which test(s) we should use for this ... I guess we'd want a fixed seed ... and things like concurrent merges could mess up the numbers if the threads schedule differently / IO gets slower ... hrmm. Mike McCandless http://blog.mikemccandless.com --------------------------------------------------------------------- To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org For additional commands, e-mail: dev-h...@lucene.apache.org
