Many thanks for the analysis !
I'm going to test with 4K on heavy mssql database to see if I'm seeing improvement on ios/latency. I'll report results in this thread. ----- Mail original ----- De: "Trent Lloyd" <[email protected]> À: "ceph-users" <[email protected]> Envoyé: Vendredi 10 Mai 2019 09:59:39 Objet: [ceph-users] Poor performance for 512b aligned "partial" writes from Windows guests in OpenStack + potential fix I recently was investigating a performance problem for a reasonably sized OpenStack deployment having around 220 OSDs (3.5" 7200 RPM SAS HDD) with NVMe Journals. The primary workload is Windows guests backed by Cinder RBD volumes. This specific deployment is Ceph Jewel (FileStore + SimpleMessenger) which while it is EOL, the issue is reproducible on current versions and also on BlueStore however for different reasons than FileStore. Generally the Ceph cluster was suffering from very poor outlier performance, the numbers change a little bit depending on the exact situation but roughly 80% of I/O was happening in a "reasonable" time of 0-200ms but 5-20% of I/O operations were taking excessively long anywhere from 500ms through to 10-20+ seconds. However the normal metrics for commit and apply latency were normal, and in fact, this latency was hard to spot in the performance metrics available in jewel. Previously I more simply considered FileStore to have the "commit" (to journal) stage where it was written to the journal and it is OK to return to the client and then the "apply" (to disk) stage where it was flushed to disk and confirmed so that the data could be purged from the journal. However there is really a third stage in the middle where FileStore submits the I/O to the operating system and this is done before the lock on the object is released. Until that succeeds another operation cannot write to the same object (generally being a 4MB area of the disk). I found that the fstore_op threads would get stuck for hundreds of MS or more inside of pwritev() which was blocking inside of the kernel. Normally we expect pwritev() to be buffered I/O into the page cache and return quite fast however in this case the kernel was in a few percent of cases blocking with the stack trace included at the end of the e-mail [1]. My finding from that stack is that inside __block_write_begin_int we see a call to out_of_line_wait_on_bit call which is really an inlined call for wait_on_buffer which occurs in linux/fs/buffer.c in the section around line 2000-2024 with the comment "If we issued read requests - let them complete." ( [ https://github.com/torvalds/linux/blob/a2d635decbfa9c1e4ae15cb05b68b2559f7f827c/fs/buffer.c#L2002 | https://github.com/torvalds/linux/blob/a2d635decbfa9c1e4ae15cb05b68b2559f7f827c/fs/buffer.c#L2002 ] ) My interpretation of that code is that for Linux to store a write in the page cache, it has to have the entire 4K page as that is the granularity of which it tracks the dirty state and it needs the entire 4K page to later submit back to the disk. Since we wrote a part of the page, and the page wasn't already in the cache, it has to fetch the remainder of the page from the disk. When this happens, it blocks waiting for this read to complete before returning from the pwritev() call - hence our normally buffered write blocks. This holds up the tp_fstore_op thread, of which there are (by default) only 2-4 such threads trying to process several hundred operations per second. Additionally the size of the osd_op_queue is bounded, and operations do not clear out of this queue until the tp_fstore_op thread is done. Which ultimately means that not only are these partial writes delayed but it knocks on to delay other writes behind them because of the constrained thread pools. What was further confusing to this, is that I could easily reproduce this in a test deployment using an rbd benchmark that was only writing to a total disk size of 256MB which I would easily have expected to fit in the page cache: rbd create -p rbd --size=256M bench2 rbd bench-write -p rbd bench2 --io-size 512 --io-threads 256 --io-total 256M --io-pattern rand This is explained by the fact that on secondary OSDs (at least, there was some refactoring of fadvise which I have not fully understood as of yet), FileStore is using fadvise FADVISE_DONTNEED on the objects after write which causes the kernel to immediately discard them from the page cache without any regard to their statistics of being recently/frequently used. The motivation for this addition appears to be that on a secondary OSD we don't service reads (only writes) and so therefor we can optimize memory usage by throwing away this object and in theory leaving more room in the page cache for objects which we are primary for and expect to actually service reads from a client for. Unfortunately this behavior does not take into account partial writes, where we now pathologically throw away the cached copy instantly such that a write even 1 second later will have to fetch the page from disk again. I also found that this FADVISE_DONTNEED is issue not only during filestore sync but also by the WBThrottle - which as this cluster was quite busy was constantly flushing writes leading to the cache being discarded almost instantly. Changing filestore_fadvise to False on this cluster lead to a significant performance increase as it could now cache the pages in memory in many cases. The number of reads from disk was reduced from around 40/second to 2/second, and the number of slow writes (>200ms) operations was reduced by 75%. I wrote a script to parse ceph-osd logs with debug_filestore=10 or 15 to report the time spent inside of write() as well as to count and report on the number of operations that are unaligned and also slow. It's a bit rough but you can find it here: [ https://github.com/lathiat/ceph-tools/blob/master/fstore_op_latency.rb | https://github.com/lathiat/ceph-tools/blob/master/fstore_op_latency.rb ] It does not solve the problem entirely, in that a filestore thread can still be blocked in such a case where it is not cached - but the pathological case of never having it in the cache is removed at least. Understanding this problem, I looked to the situation for BlueStore. BlueStore suffers from a similar issue in that the performance is quite poor due to both fadvise and also because it is check-summing the data in 4k blocks so needs to read the rest of the block in, despite not having the limitations of the Linux page cache to deal with. I have not yet further fully investigated BlueStore implementation other than to note the following doc talking about how such writes are handled and a possible future improvement to submit partial writes into the WAL before reading the rest of the block, which is apparently not done currently (and would be a great optimization): [ http://docs.ceph.com/docs/mimic/dev/bluestore/ | http://docs.ceph.com/docs/mimic/dev/bluestore/ ] Moving onto a full solution for this issue. We can tell Windows guests to send 4k-aligned I/O where possible by setting the physical_block_size hint on the disk. This support was added mainly for the incoming new series of hard drives which also have 4k blocks internally, and also need to do a similar 'read-modify-update' operation in the case where a smaller write is done. In this case Windows tries to align the I/O to 4k as much as possible, at the most basic level for example when a new file is created, it will pad out the write to the block to the nearest 4k. You can read more about support for that here: [ https://support.microsoft.com/en-au/help/2510009/microsoft-support-policy-for-4k-sector-hard-drives-in-windows | https://support.microsoft.com/en-au/help/2510009/microsoft-support-policy-for-4k-sector-hard-drives-in-windows ] On a basic test, booting a Windows 2016 instance and then installing several months of Windows Updates the number of partial writes was reduced from 23% (753090 / 3229597) to 1.8% (54535 / 2880217) - many of which were during early boot and don't re-occur once the VM is running. I have submitted a patch to the OpenStack Cinder RBD driver to support setting this parameter. You can find that here: [ https://review.opendev.org/#/c/658283/ | https://review.opendev.org/#/c/658283/ ] I did not have much luck finding information about any of this online when I searched, so this e-mail is serving largely to document my findings for others. But I am also looking for input from anyone as to anything I have missed, confirming my analysis as sound, review for my Cinder patch, etc. There is also likely scope to make this same patch to report a physical_block_size=4096 on other Ceph consumers such as the new(ish) iSCSI gateway, etc. Regards, Trent [1] fstore_op pwritev blocking stack trace - if anyone is interested in the perf data, flamegraph, etc - I'd be happy to share. tp_fstore_op ceph::buffer::list::write_fd pwritev64 entry_SYSCALL_64_after_hwframe do_syscall_64 sys_pwritev do_pwritev vfs_writev do_iter_write do_iter_readv_writev xfs_file_write_iter xfs_file_buffered_aio_write iomap_file_buffered_write iomap_apply iomap_write_actor iomap_write_begin.constprop.18 __block_write_begin_int out_of_line_wait_on_bit __wait_on_bit bit_wait_io io_schedule schedule __schedule finish_task_switch _______________________________________________ ceph-users mailing list [email protected] http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com _______________________________________________ ceph-users mailing list [email protected] http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
