In my measurements 0.94 has been getting faster with each release in both read and write performance. I wonder how representative PE is after all; it only tests via the local FS layer (not HDFS), among other issues.
-- Lars ----- Original Message ----- From: Jean-Marc Spaggiari <[email protected]> To: [email protected]; lars hofhansl <[email protected]> Cc: Sent: Friday, June 28, 2013 8:03 PM Subject: Re: 30% random performance in 0.95+ I think we should do that on 0.94 as well. I don't see any good reason to not do it. JM 2013/6/28 lars hofhansl <[email protected]>: > Yep. > Now the question is: Make these changes to 0.94 as well? Or just document > these better. > > -- Lars > > > > ----- Original Message ----- > From: Andrew Purtell <[email protected]> > To: "[email protected]" <[email protected]>; lars hofhansl > <[email protected]> > Cc: > Sent: Friday, June 28, 2013 2:08 PM > Subject: Re: 30% random performance in 0.95+ > > I've been thinking about how to periodically search through some of our > parameter space to see what changes to defaults are better all the way > around. Probably will so something based on Bigtop. > > > On Friday, June 28, 2013, lars hofhansl wrote: > >> And indeed just this makes a tremendous difference. Unpatched 0.94 with >> 40% block cache configured is actually faster than 0.95 with the same block >> cache size. >> >> -- Lars >> >> >> >> ----- Original Message ----- >> From: lars hofhansl <[email protected] <javascript:;>> >> To: "[email protected] <javascript:;>" <[email protected]<javascript:;> >> > >> Cc: >> Sent: Friday, June 28, 2013 1:34 PM >> Subject: Re: 30% random performance in 0.95+ >> >> Thanks JM, >> >> HBASE-8450 (r1485562) is interesting. It increases (among other things) >> the block cache percentage from 24 to 40%, which would lead to a higher >> probability of a future random read to hit an already cached block. >> >> >> -- Lars >> >> >> >> ----- Original Message ----- >> From: Jean-Marc Spaggiari <[email protected] <javascript:;>> >> To: [email protected] <javascript:;>; lars hofhansl >> <[email protected]<javascript:;> >> > >> Cc: >> Sent: Friday, June 28, 2013 1:18 PM >> Subject: Re: 30% random performance in 0.95+ >> >> I have the script done to run over a list of "svn releases", so if >> required, just give me a bunch of them or a range and I can restart. >> Just keep me posted. >> >> JM >> >> 2013/6/28 lars hofhansl <[email protected] <javascript:;>>: >> > I did a few more test (on my laptop, which is not quite representative), >> and found only a 2-3% improvement from HBASE-8001+HBASE-8012 in the end. >> > I'll look through the issues that you identified. >> > >> > -- Lars >> > >> > >> > >> > ----- Original Message ----- >> > From: Jean-Marc Spaggiari <[email protected] <javascript:;>> >> > To: [email protected] <javascript:;> >> > Cc: >> > Sent: Friday, June 28, 2013 12:51 PM >> > Subject: Re: 30% random performance in 0.95+ >> > >> > Sorry folks, >> > >> > I'm a bit late to run the tests... 0.94.8 and 0.94.9 are currently >> > running, but here is what I have been able to capture so far for 0.95 >> > over the last year: >> > r1357480 1513196 >> > r1367009 1440244.4 >> > r1375812 1287143.5 >> > r1381671 1287200.2 >> > r1388620 1295262.6 >> > r1394335 1022140.2 >> > r1403898 884171.9 >> > r1410631 804229.9 >> > r1419787 846816.9 >> > r1426557 853535.3 >> > r1433514 873265.1 >> > r1438972 840666.9 >> > r1446106 877432.2 >> > r1452661 883974.8 >> > r1458421 882233.3 >> > r1464267 847000.8 >> > r1478964 877433.5 >> > r1485868 744905.5 >> > r1494869 765105.9 >> > >> > So seems that there was some improvements between r1367009 and >> > r1403898 but they are old. Also another major improvement between >> > r1478964 and r1485868... >> > >> > Let me know if you want me to dig further and I will be very happy to do >> so. >> > >> > JM >> > >> > 2013/6/28 Stack <[email protected] <javascript:;>>: >> >> On Fri, Jun 28, 2013 at 10:53 AM, lars hofhansl >> >> <[email protected]<javascript:;>> >> wrote: >> >> >> >>> I partially tracked this down to HBASE-8001 and HBASE-8012 by looking >> at >> >>> the call stacks in a profiling session. >> >>> HBASE-8767 is a backport of both patched to 0.94. >> >>> >> >> >> >> Sounds like nice work by Raymond Liu... >> >> St.Ack >> > >> > > > -- > Best regards, > > - Andy > > Problems worthy of attack prove their worth by hitting back. - Piet Hein > (via Tom White) >
