On Mon, Aug 20 2018, Ulrich Windl wrote: >>>> Jeff King <peff@xxxxxxxx> schrieb am 16.08.2018 um 22:55 in Nachricht > <20180816205556.GA8257@xxxxxxxxxxxxxxxxxxxxx>: >> On Thu, Aug 16, 2018 at 10:35:53PM +0200, Ævar Arnfjörð Bjarmason wrote: >> >>> This is all interesting, but I think unrelated to what Ulrich is talking >>> about. Quote: >>> >>> Between the two phases of "git fsck" (checking directories and >>> checking objects) there was a break of several seconds where no >>> progress was indicated >>> >>> I.e. it's not about the pause you get with your testcase (which is >>> certainly another issue) but the break between the two progress bars. >> >> I think he's talking about both. What I said responds to this: > > Hi guys! > > Yes, I was wondering what git does between the two visible phases, and between > the lines I was suggesting another progress message between those phases. At > least the maximum unspecific three-dot-message "Thinking..." could be displayed > ;-) Of course anything more appropriate would be welcome. > Also that message should only be displayed if it's foreseeable that the > operation will take significant time. In my case (I just repeated it a few > minutes ago) the delay is significant (at least 10 seconds). As noted earlier I > was hoping to capture the timing in a screencast, but it seems all the delays > were just optimized away in the recording. > >> >>> >> During "git gc" the writing objects phase did not update for some >>> >> seconds, but then the percentage counter jumped like from 15% to 42%. >> >> But yeah, I missed that the fsck thing was specifically about a break >> between two meters. That's a separate problem, but also worth >> discussing (and hopefully much easier to address). >> >>> If you fsck this repository it'll take around (on my spinning rust >>> server) 30 seconds between 100% of "Checking object directories" before >>> you get any output from "Checking objects". >>> >>> The breakdown of that is (this is from approximate eyeballing): >>> >>> * We spend 1-3 seconds just on this: >>> >> > https://github.com/git/git/blob/63749b2dea5d1501ff85bab7b8a7f64911d21dea/pack > >> -check.c#L181 >> >> OK, so that's checking the sha1 over the .idx file. We could put a meter >> on that. I wouldn't expect it to generally be all that slow outside of >> pathological cases, since it scales with the number of objects (and 1s >> is our minimum update anyway, so that might be OK as-is). Your case has >> 13M objects, which is quite large. > > Sometimes an oldish CPU could bring performance surprises, maybe. Anyway my > CPU is question is an AMD Phenom2 quad-core with 3.2GHz nominal, and there is a > classic spinning disk with 5400RPM built in... > >> >>> * We spend the majority of the ~30s on this: >>> >> > https://github.com/git/git/blob/63749b2dea5d1501ff85bab7b8a7f64911d21dea/pack > >> -check.c#L70-L79 >> >> This is hashing the actual packfile. This is potentially quite long, >> especially if you have a ton of big objects. > > That seems to apply. BTW: Is there a way go get some repository statistics > like a histogram of object sizes (or whatever that might be useful to help > making decisions)? The git-sizer program is really helpful in this regard: https://github.com/github/git-sizer >> >> I wonder if we need to do this as a separate step anyway, though. Our >> verification is based on index-pack these days, which means it's going >> to walk over the whole content as part of the "Indexing objects" step to >> expand base objects and mark deltas for later. Could we feed this hash >> as part of that walk over the data? It's not going to save us 30s, but >> it's likely to be more efficient. And it would fold the effort naturally >> into the existing progress meter. >> >>> * Wes spend another 3-5 seconds on this QSORT: >>> >> > https://github.com/git/git/blob/63749b2dea5d1501ff85bab7b8a7f64911d21dea/pack > >> -check.c#L105 >> >> That's a tough one. I'm not sure how we'd count it (how many compares we >> do?). And each item is doing so little work that hitting the progress >> code may make things noticeably slower. > > If it's sorting, maybe add some code like (wild guess): > > if (objects_to_sort > magic_number) > message("Sorting something..."); I think a good solution to these cases is to just introduce something to the progress.c mode where it learns how to display a counter where we don't know what the end-state will be. Something like your proposed magic_number can just be covered under the more general case where we don't show the progress bar unless it's been 1 second (which I believe is the default). >> >> Again, your case is pretty big. Just based on the number of objects, >> linux.git should be 1.5-2.5 seconds on your machine for the same >> operation. Which I think may be small enough to ignore (or even just >> print a generic before/after). It's really the 30s packfile hash that's >> making the whole thing so terrible. >> >> -Peff
