Hi John, I ran my 90k query set against the broken version of last Friday which produced a grand total of 1426117 hits (the sum of hits of each queries). The fixed version yields a total 1633307, which is not so different. For the number of bytes, this is bit tougher to evaluate, not sure how i can do this.
Thanks, -----Original Message----- From: K. John Wu [mailto:[email protected]] Sent: Friday, March 16, 2012 3:42 PM To: Dominique Prunier Cc: FastBit Users Subject: Re: [FastBit-users] PATCH: new CS_PATTERN_MATCH define added to match LIKE patterns case-sensitively and perform specific optimizations Hi, Dominique, Thanks for the info. There is one piece of information that might be able to resolve the mystery about the timing difference. How different are the number of answers you get back from the previous version of the code (incorrect) compared to the correct answers? If the correct answers are much more than reported previously, then simply by the fact there are more answers, it would take longer to process the bitmaps. The more direct measure is the number of bytes (of memory) used by the bit vector objects. Presumably, that information is a little harder to gather. John On 3/16/12 12:30 PM, Dominique Prunier wrote: > Hi John, > > I just tried the new code, it seems that it saves a few percents (1-5% > depending how much pattern matching there is in the query) but we're still > far from the initial results. > > Thanks, > > -----Original Message----- > From: K. John Wu [mailto:[email protected]] > Sent: Friday, March 16, 2012 2:06 PM > To: Dominique Prunier > Cc: FastBit Users > Subject: Re: [FastBit-users] PATCH: new CS_PATTERN_MATCH define added to > match LIKE patterns case-sensitively and perform specific optimizations > > Hi, Dominique, > > The function ibis::category::fillIndex now sorts the dictionary before > actually creating the index. The source code is SVN 492. You could > give it a try and see if it helps. > > John > > > On 3/16/12 6:34 AM, Dominique Prunier wrote: >> Hey John, >> >> I tried the sort. It ends up speeding up some cases but slowing down some >> others (i guess when there are many values, you are paying the sort). At the >> end, on my large query test set, it ended up being slower so i reverted it. >> >> Like i said earlier, this exact same set is now 6 times slower. I think it >> is related to the fact that my bit vectors are much more complex (and more >> distributed) and more likely to match something (since the previous results >> were pure garbage), but i'm not sure why it make such a difference (i'd >> assume the first N one would be as dispersed that an set of N vectors). The >> hot spot is ibis::bitvector::do_cnt and i have the impression that it is >> called quite often with a test like cnt() != 0 or > 0 which could probably >> be simplified. I'll take a look at that if i have time today. >> >> Thanks, >> >> -----Original Message----- >> From: [email protected] >> [mailto:[email protected]] On Behalf Of Dominique Prunier >> Sent: Thursday, March 15, 2012 5:08 PM >> To: K. John Wu >> Cc: FastBit Users >> Subject: Re: [FastBit-users] PATCH: new CS_PATTERN_MATCH define added to >> match LIKE patterns case-sensitively and perform specific optimizations >> >> Hi John, >> >> About the perf, i'm trying to understand what could justify such a >> difference (my ~90k queries test jumped from 20s to 120s while we should be >> handling the same number of bit vectors). So far, it seems that ~70% of the >> time is spent in ibis::bitvector::do_cnt(). I'll try the sort trick to see >> how it affects performance. I tend to be carefull with profiler results and >> favor your knowledge of the codebase. >> >> Thanks, >> >> -----Original Message----- >> From: K. John Wu [mailto:[email protected]] >> Sent: Thursday, March 15, 2012 4:56 PM >> To: Dominique Prunier >> Cc: FastBit Users >> Subject: Re: [FastBit-users] PATCH: new CS_PATTERN_MATCH define added to >> match LIKE patterns case-sensitively and perform specific optimizations >> >> Hi, Dominique, >> >> Regarding the test cases, whenever you are ready to share is fine. >> >> Regarding the performance, my guess is that the bit vectors are more >> spread out in the new corrected approach. When the bit vectors are >> spread out, FastBit needs to read more bytes in order to get them out. >> If this is indeed the case, the thing to do might to sort the values >> before returning them from ibis::dictionary::patternSearch with >> something like >> >> std::sort(matches.begin(), matches.end()); >> >> John >> >> >> On 3/15/12 12:24 PM, Dominique Prunier wrote: >>> Hey John, >>> >>> The problem is that right know, my test suite mainly tests the >>> import and not so much the queries (i have maybe 10-20 test cases >>> that actually evaluates queries). It has been basically modeled >>> after our existing test suite for MySQL/Oracle and is supposed to >>> prove that the imported data is equivalent to our source. >>> >>> My next step is to do a micro testing framework to test queries. I >>> have a very simplistic example which only uses CATEGORY and LONG >>> (the only two types that we're using), but it is not convenient >>> enough to use. >>> >>> Be sure that i'll post in the group if i have something good enough >>> to be shared :) >>> >>> Regarding the bug i found yesterday, surprisingly enough, fixing it >>> changed the performance quite significantly. I'm still not sure why >>> (the number of selected bitmap was the same), but i'll try to >>> investigate a bit further. >>> >>> Thanks, >>> >> _______________________________________________ >> FastBit-users mailing list >> [email protected] >> https://hpcrdm.lbl.gov/cgi-bin/mailman/listinfo/fastbit-users _______________________________________________ FastBit-users mailing list [email protected] https://hpcrdm.lbl.gov/cgi-bin/mailman/listinfo/fastbit-users
