From: Trond Myklebust Subject: Re: [PATCH v2] flow control for WRITE requests Date: Tue, 02 Jun 2009 18:12:15 -0400 Message-ID: <1243980736.4868.314.camel@heimdal.trondhjem.org> References: <49C93526.70303@redhat.com> <20090324211917.GJ19389@fieldses.org> <4A1D9210.8070102@redhat.com> <1243457149.8522.68.camel@heimdal.trondhjem.org> <4A1EB09A.8030809@redhat.com> <1243892886.4868.74.camel@heimdal.trondhjem.org> <4A257167.9090304@redhat.com> Mime-Version: 1.0 Content-Type: text/plain Cc: "J. Bruce Fields" , NFS list To: Peter Staubach Return-path: Received: from mail-out2.uio.no ([129.240.10.58]:46496 "EHLO mail-out2.uio.no" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1754866AbZFBWMT (ORCPT ); Tue, 2 Jun 2009 18:12:19 -0400 In-Reply-To: <4A257167.9090304@redhat.com> Sender: linux-nfs-owner@vger.kernel.org List-ID: On Tue, 2009-06-02 at 14:37 -0400, Peter Staubach wrote: > Trond Myklebust wrote: > > > > So, how about doing this by modifying balance_dirty_pages() instead? > > Limiting pages on a per-inode basis isn't going to solve the common > > problem of 'ls -l' performance, where you have to stat a whole bunch of > > files, all of which may be dirty. To deal with that case, you really > > need an absolute limit on the number of dirty pages. > > > > Currently, we have only relative limits: a given bdi is allowed a > > maximum percentage value of the total write back cache size... We could > > add a 'max_pages' field, that specifies an absolute limit at which the > > vfs should start writeback. > > Interesting thought. From a high level, it sounds like a good > strategy. The details start to get a little troubling to me > though. > > First thing that strikes me is that this may result in > suboptimal WRITE requests being issued over the wire. If the > page quota is filled with many pages from one file and just a > few from another due to timing, we may end up issuing small > over the wire WRITE requests for the one file, even during > normal operations. balance_dirty_pages() will currently call writeback_inodes() to actually flush out the pages. The latter will again check the super block dirty list to determine candidate files; it doesn't favour the particular file on which we called balance_dirty_pages_ratelimited(). That said, balance_dirty_pages_ratelimited() does take the mapping as an argument. You could, therefore, in theory have it make decisions on a per-mapping basis. > We don't want to flush pages in the page cache until an entire > wsize'd transfer can be constructed for the specific file. > Thus, it seems to me that we still need to track the number of > dirty pages per file. > > We also need to know that those pages are contiguous in the > file. We can determine, heuristically, whether the pages are > contiguous in the file or not by tracking the access pattern. > For random access, we can assume that the pages are not > contiguous and we can assume that they are contiguous for > sequential access. This isn't perfect and can be fooled, > but should hold for most applications which access files > sequentially. > > Also, we don't want to proactively flush the cache if the > application is doing random access. The application may come > back to the page and we could get away with a single WRITE > instead of multiple WRITE requests for the same page. With > sequential access, we can generally know that it is safe to > proactively flush pages because the application won't be > accessing them again. Once again, this heuristic is not > foolproof, but holds most of the time. I'm not sure I follow you here. Why is the random access case any different to the sequential access case? Random writes are obviously a pain to deal with since you cannot predict access patterns. However, AFAICS if we want to provide a faster generic stat(), then we need to deal with random writes too: a gigabyte of data will take even longer to flush out when it is in the form of non-contiguous writes. > For the ls case, we really want to manage the page cache on a > per-directory of files case. I don't think that this is going > to happen. The only directions to go from there are more > coarse, per-bdi, or less coarse, per-file. Ugh. No... > If we go the per-bdi approach, then we would need to stop > all modifications to the page cache for that particular bdi > during the duration of the ls processing. Otherwise, as we > stat 1 file at a time, the other files still needing to be > stat'd would just refill the page cache with dirty pages. > We could solve this by setting the max_pages limit to be a > reasonable number to flush per file, but then that would be > too small a limit for the entire file system. True, but if you have applications writing to all the files in your directory, then 'ls -l' performance is likely to suck anyway. Even if you do have per-file limits, those write-backs to the other files will be competing for RPC slots with the write-backs from the file that is being stat()ed. > So, I don't see how to get around managing the page cache on > a per-file basis, at least to some extent, in order to manage > the amount of dirty data that must be flushed. > > It does seem like the right way to do this is via a combination > of per-bdi and per-file support, but I am not sure that we have > the right information at the right levels to achieve this now. > > Thanx... > > ps In the long run, I'd like to see us merge something like the fstatat() patches that were hacked together at the LSF'09 conference. If applications can actually tell the NFS client that they don't care about a/c/mtime accuracy, then we can avoid this whole flushing nonsense altogether. It would suffice to teach 'ls' to start using the AT_NO_TIMES flag that we defined... Cheers Trond