MMapDirectory really should be (is supposed to be) better than SimpleFSDirectory for your usage case.
Memory mapped pages do not have to fit into your 64 GB physical space, but the "hot" pages (parts of the index that you are actively querying) ideally would fit mostly in free RAM on your box to have OK search performance. Run with as small a JVM heap as possible so the OS has the most RAM to keep such pages hot. Since you are getting OK performance with SimpleFSDirectory it sounds like you do have enough free RAM for the parts of the index you are searching... But, I believe you (system locks up with MMapDirectory for you use-case), so there is a bug somewhere! And I wish we could get to the bottom of that, and fix it. Also, this (system locks up when using MMapDirectory) sounds different from the "Lucene fsyncs files that it doesn't need to" bug, right? Mike McCandless http://blog.mikemccandless.com On Mon, Mar 15, 2021 at 4:28 PM Rahul Goswami <rahul196...@gmail.com> wrote: > Uwe, > I understand that mmap would only map *a part* of the index from virtual > address space to physical memory as and when the pages are requested. > However the limitation on our side is that in most cases, we cannot ask for > more than 128 GB RAM (and unfortunately even that would be a stretch) for > the Solr machine. > > I have read and re-read the article you referenced in the past :) It's > brilliantly written and did help clarify quite a few things for me I must > say. However, at the end of the day, there is only so much the OS (at least > Windows) can do before it starts to swap different pages in a 2-3 TB index > into 64 GB of physical space, isn't that right ? The CPU usage spikes to > 100% at such times and the machine becomes totally unresponsive. Turning on > SimpleFSDIrectory at such times does rid us of this issue. I understand > that we are losing out on performance by an order of magnitude compared to > mmap, but I don't know any alternate solution. Also, since most of our use > cases are more write-heavy than read-heavy, we can afford to compromise on > the search performance due to SimpleFS. > > Please let me know still, if there is anything about my explanation that > doesn't sound right to you. > > Thanks, > Rahul > > On Mon, Mar 15, 2021 at 3:54 PM Uwe Schindler <u...@thetaphi.de> wrote: > >> This is not true. Memory mapping does not need to load the index into >> ram, so you don't need so much physical memory. Paging is done only between >> index files and ram, that's what memory mapping is about. >> >> Please read the blog post: >> https://blog.thetaphi.de/2012/07/use-lucenes-mmapdirectory-on-64bit.html >> >> Uwe >> >> Am March 15, 2021 7:43:29 PM UTC schrieb Rahul Goswami < >> rahul196...@gmail.com>: >>> >>> Mike, >>> Yes I am using a 64 bit JVM on Windows. I haven't tried reproducing the >>> issue on Linux yet. In the past we have had problems with mmap on Windows >>> with the machine freezing. The rationale I gave to myself is the amount of >>> disk and CPU activity for paging in and out must be intense for the OS >>> while trying to map an index that large into 64 GB of heap. Also since it's >>> an on-premise deployment, we can't expect the customers of the product to >>> provide nodes with > 400 GB RAM which is what *I think* would be required >>> to get a decent performance with mmap. Hence we had to switch to >>> SimpleFSDirectory. >>> >>> As for the fsync behavior, you are right. I tried with >>> NRTCachingDirectoryFactory as well which defaults to using mmap underneath >>> and still makes fsync calls for already existing index files. >>> >>> Thanks, >>> Rahul >>> >>> On Mon, Mar 15, 2021 at 3:15 PM Michael McCandless < >>> luc...@mikemccandless.com> wrote: >>> >>>> Thanks Rahul. >>>> >>>> > primary reason being that memory mapping multi-terabyte indexes is >>>> not feasible through mmap >>>> >>>> Hmm, that is interesting -- are you using a 64 bit JVM? If so, what >>>> goes wrong with such large maps? Lucene's MMapDirectory should chunk the >>>> mapping to deal with ByteBuffer int only address space. >>>> >>>> SimpleFSDirectory usually has substantially worse performance than >>>> MMapDirectory. >>>> >>>> Still, I suspect you would hit the same issue if you used other >>>> FSDirectory implementations -- the fsync behavior should be the same. >>>> >>>> Mike McCandless >>>> >>>> http://blog.mikemccandless.com >>>> >>>> >>>> On Fri, Mar 12, 2021 at 1:46 PM Rahul Goswami <rahul196...@gmail.com> >>>> wrote: >>>> >>>>> Thanks Michael. For your question...yes I am running Solr on Windows >>>>> and running it with SimpleFSDirectoryFactory (primary reason being that >>>>> memory mapping multi-terabyte indexes is not feasible through mmap). I >>>>> will >>>>> create a Jira later today with the details in this thread and assign it to >>>>> myself. Will take a shot at the fix. >>>>> >>>>> Thanks, >>>>> Rahul >>>>> >>>>> On Fri, Mar 12, 2021 at 10:00 AM Michael McCandless < >>>>> luc...@mikemccandless.com> wrote: >>>>> >>>>>> I think long ago we used to track which files were actually dirty (we >>>>>> had written bytes to) and only fsync those ones. But something went >>>>>> wrong >>>>>> with that, and at some point we "simplified" this logic, I think on the >>>>>> assumption that asking the OS to fsync a file that does in fact exist yet >>>>>> indeed has not changed would be harmless? But somehow it is not in your >>>>>> case? Are you on Windows? >>>>>> >>>>>> I tried to do a bit of digital archaeology and remember what >>>>>> happened here, and I came across this relevant looking issue: >>>>>> https://issues.apache.org/jira/browse/LUCENE-2328. That issue moved >>>>>> tracking of which files have been written but not yet fsync'd down from >>>>>> IndexWriter into FSDirectory. >>>>>> >>>>>> But there was another change that then removed staleFiles from >>>>>> FSDirectory entirely.... still trying to find that. Aha, found it! >>>>>> https://issues.apache.org/jira/browse/LUCENE-6150. Phew Uwe was >>>>>> really quite upset in that issue ;) >>>>>> >>>>>> I also came across this delightful related issue, showing how a >>>>>> massive hurricane (Irene) can lead to finding and fixing a bug in Lucene! >>>>>> https://issues.apache.org/jira/browse/LUCENE-3418 >>>>>> >>>>>> > The assumption is that while the commit point is saved, no changes >>>>>> happen to the segment files in the saved generation. >>>>>> >>>>>> This assumption should really be true. Lucene writes the files, >>>>>> append only, once, and then never changes them, once they are closed. >>>>>> Pulling a commit point from Solr should further ensure that, even as >>>>>> indexing continues and new segments are written, the old segments >>>>>> referenced in that commit point will not be deleted. But apparently this >>>>>> "harmless fsync" Lucene is doing is not so harmless in your use case. >>>>>> Maybe open an issue and pull out the details from this discussion onto >>>>>> it? >>>>>> >>>>>> Mike McCandless >>>>>> >>>>>> http://blog.mikemccandless.com >>>>>> >>>>>> >>>>>> On Fri, Mar 12, 2021 at 9:03 AM Michael Sokolov <msoko...@gmail.com> >>>>>> wrote: >>>>>> >>>>>>> Also - I should have said - I think the first step here is to write a >>>>>>> focused unit test that demonstrates the existence of the extra fsyncs >>>>>>> that we want to eliminate. It would be awesome if you were able to >>>>>>> create such a thing. >>>>>>> >>>>>>> On Fri, Mar 12, 2021 at 9:00 AM Michael Sokolov <msoko...@gmail.com> >>>>>>> wrote: >>>>>>> > >>>>>>> > Yes, please go ahead and open an issue. TBH I'm not sure why this >>>>>>> is >>>>>>> > happening - there may be a good reason?? But let's explore it >>>>>>> using an >>>>>>> > issue, thanks. >>>>>>> > >>>>>>> > On Fri, Mar 12, 2021 at 12:16 AM Rahul Goswami < >>>>>>> rahul196...@gmail.com> wrote: >>>>>>> > > >>>>>>> > > I can create a Jira and assign it to myself if that's ok (?). I >>>>>>> think this can help improve commit performance. >>>>>>> > > Also, to answer your question, we have indexes sometimes going >>>>>>> into multiple terabytes. Using the replication handler for backup would >>>>>>> mean requiring a disk capacity more than 2x the index size on the >>>>>>> machine >>>>>>> at all times, which might not be feasible. So we directly back the >>>>>>> index up >>>>>>> from the Solr node to a remote repository. >>>>>>> > > >>>>>>> > > Thanks, >>>>>>> > > Rahul >>>>>>> > > >>>>>>> > > On Thu, Mar 11, 2021 at 4:09 PM Michael Sokolov < >>>>>>> msoko...@gmail.com> wrote: >>>>>>> > >> >>>>>>> > >> Well, it certainly doesn't seem necessary to fsync files that >>>>>>> are >>>>>>> > >> unchanged and have already been fsync'ed. Maybe there's an >>>>>>> opportunity >>>>>>> > >> to improve it? On the other hand, support for external processes >>>>>>> > >> reading Lucene index files isn't likely to become a feature of >>>>>>> Lucene. >>>>>>> > >> You might want to consider using Solr replication to power your >>>>>>> > >> backup? >>>>>>> > >> >>>>>>> > >> On Thu, Mar 11, 2021 at 2:52 PM Rahul Goswami < >>>>>>> rahul196...@gmail.com> wrote: >>>>>>> > >> > >>>>>>> > >> > Thanks Michael. I thought since this discussion is closer to >>>>>>> the code than most discussions on the solr-users list, it seemed like a >>>>>>> more appropriate forum. Will be mindful going forward. >>>>>>> > >> > On your point about new segments, I attached a debugger and >>>>>>> tried to do a new commit (just pure Solr commit, no backup process >>>>>>> running), and the code indeed does fsync on a pre-existing segment file. >>>>>>> Hence I was a bit baffled since it challenged my fundamental >>>>>>> understanding >>>>>>> that segment files once written are immutable, no matter what (unless >>>>>>> picked up for a merge of course). Hence I thought of reaching out, in >>>>>>> case >>>>>>> there are scenarios where this might happen which I might be unaware of. >>>>>>> > >> > >>>>>>> > >> > Thanks, >>>>>>> > >> > Rahul >>>>>>> > >> > >>>>>>> > >> > On Thu, Mar 11, 2021 at 2:38 PM Michael Sokolov < >>>>>>> msoko...@gmail.com> wrote: >>>>>>> > >> >> >>>>>>> > >> >> This isn't a support forum; solr-users@ might be more >>>>>>> appropriate. On >>>>>>> > >> >> that list someone might have a better idea about how the >>>>>>> replication >>>>>>> > >> >> handler gets its list of files. This would be a good list to >>>>>>> try if >>>>>>> > >> >> you wanted to propose a fix for the problem you're having. >>>>>>> But since >>>>>>> > >> >> you're here -- it looks to me as if IndexWriter indeed syncs >>>>>>> all "new" >>>>>>> > >> >> files in the current segments being committed; look in >>>>>>> > >> >> IndexWriter.startCommit and SegmentInfos.files. Caveat: (1) >>>>>>> I'm >>>>>>> > >> >> looking at this code for the first time, and (2) things may >>>>>>> have been >>>>>>> > >> >> different in 7.7.2? Sorry I don't know for sure, but are you >>>>>>> sure that >>>>>>> > >> >> your backup process is not attempting to copy one of the new >>>>>>> files? >>>>>>> > >> >> >>>>>>> > >> >> On Thu, Mar 11, 2021 at 1:35 PM Rahul Goswami < >>>>>>> rahul196...@gmail.com> wrote: >>>>>>> > >> >> > >>>>>>> > >> >> > Hello, >>>>>>> > >> >> > Just wanted to follow up one more time to see if this is >>>>>>> the right form for my question? Or is this suitable for some other >>>>>>> mailing >>>>>>> list? >>>>>>> > >> >> > >>>>>>> > >> >> > Best, >>>>>>> > >> >> > Rahul >>>>>>> > >> >> > >>>>>>> > >> >> > On Sat, Mar 6, 2021 at 3:57 PM Rahul Goswami < >>>>>>> rahul196...@gmail.com> wrote: >>>>>>> > >> >> >> >>>>>>> > >> >> >> Hello everyone, >>>>>>> > >> >> >> Following up on my question in case anyone has any idea. >>>>>>> Why it's important to know this is because I am thinking of allowing the >>>>>>> backup process to not hold any lock on the index files, which should >>>>>>> allow >>>>>>> the fsync during parallel commits. BUT, in case doing an fsync on >>>>>>> existing >>>>>>> segment files in a saved commit point DOES have an effect, it might >>>>>>> render >>>>>>> the backed up index in a corrupt state. >>>>>>> > >> >> >> >>>>>>> > >> >> >> Thanks, >>>>>>> > >> >> >> Rahul >>>>>>> > >> >> >> >>>>>>> > >> >> >> On Fri, Mar 5, 2021 at 3:04 PM Rahul Goswami < >>>>>>> rahul196...@gmail.com> wrote: >>>>>>> > >> >> >>> >>>>>>> > >> >> >>> Hello, >>>>>>> > >> >> >>> We have a process which backs up the index (Solr 7.7.2) >>>>>>> on a schedule. The way we do it is we first save a commit point on the >>>>>>> index and then using Solr's /replication handler, get the list of files >>>>>>> in >>>>>>> that generation. After the backup completes, we release the commit point >>>>>>> (Please note that this is a separate backup process outside of Solr and >>>>>>> not >>>>>>> the backup command of the /replication handler) >>>>>>> > >> >> >>> The assumption is that while the commit point is saved, >>>>>>> no changes happen to the segment files in the saved generation. >>>>>>> > >> >> >>> >>>>>>> > >> >> >>> Now the issue... The backup process opens the index >>>>>>> files in a shared READ mode, preventing writes. This is causing any >>>>>>> parallel commits to fail as it seems to be complaining about the index >>>>>>> files to be locked by another process(the backup process). Upon >>>>>>> debugging, >>>>>>> I see that fsync is being called during commit on already existing >>>>>>> segment >>>>>>> files which is not expected. So, my question is, is there any reason for >>>>>>> lucene to call fsync on already existing segment files? >>>>>>> > >> >> >>> >>>>>>> > >> >> >>> The line of code I am referring to is as below: >>>>>>> > >> >> >>> try (final FileChannel file = >>>>>>> FileChannel.open(fileToSync, isDir ? StandardOpenOption.READ : >>>>>>> StandardOpenOption.WRITE)) >>>>>>> > >> >> >>> >>>>>>> > >> >> >>> in method fsync(Path fileToSync, boolean isDir) of the >>>>>>> class file >>>>>>> > >> >> >>> >>>>>>> > >> >> >>> lucene\core\src\java\org\apache\lucene\util\IOUtils.java >>>>>>> > >> >> >>> >>>>>>> > >> >> >>> Thanks, >>>>>>> > >> >> >>> Rahul >>>>>>> > >> >> >>>>>>> > >> >> >>>>>>> --------------------------------------------------------------------- >>>>>>> > >> >> To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org >>>>>>> > >> >> For additional commands, e-mail: dev-h...@lucene.apache.org >>>>>>> > >> >> >>>>>>> > >> >>>>>>> > >> >>>>>>> --------------------------------------------------------------------- >>>>>>> > >> To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org >>>>>>> > >> For additional commands, e-mail: dev-h...@lucene.apache.org >>>>>>> > >> >>>>>>> >>>>>>> --------------------------------------------------------------------- >>>>>>> To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org >>>>>>> For additional commands, e-mail: dev-h...@lucene.apache.org >>>>>>> >>>>>>> >> -- >> Uwe Schindler >> Achterdiek 19, 28357 Bremen >> https://www.thetaphi.de >> >
