Bug 203891 - Consider supporting linux' sync_file_range()
Summary: Consider supporting linux' sync_file_range()
Status: New
Alias: None
Product: Base System
Classification: Unclassified
Component: kern (show other bugs)
Version: CURRENT
Hardware: Any Any
: --- Affects Some People
Assignee: freebsd-bugs mailing list
URL:
Keywords: feature, needs-patch, performance
Depends on:
Blocks:
 
Reported: 2015-10-20 10:41 UTC by Andres Freund
Modified: 2015-10-20 19:51 UTC (History)
4 users (show)

See Also:


Attachments

Note You need to log in before you can comment on or make changes to this bug.
Description Andres Freund 2015-10-20 10:41:34 UTC
Hi,

postgresql is about to use sync_file_range(SYNC_FILE_RANGE_WRITE) to control writeback more explicitly. It'd be cool if more OSs than just linux could benefit.

Some background:
Postgres regularly 'checkpoints' it's in-memory data to disk, to be able to remove older journalling/write ahead log data. In a database with a write heavy workload that can imply a lot of writes. At the end of the checkpoint postgres then fsync()s all the files.  This unfortunately often causes latency spikes because a) the fsyncs at the end might have to write back a lot of data, unnecessarily stalling other IO b) before the fsync a lot of dirty data might accumulate kernel-side, which then also can trigger latency spikes.  Often this also leads to irregular IO with periods of no IO.

What postgres is going to do on linux is to issue sync_file_range(SYNC_FILE_RANGE_WRITE) every few (32 seems to work well) blocks during the checkpoint. That makes it rather likely that there's little dirty data remaining when the fsync()s at the end are executed, making them fast. It also prevents large amounts of dirty buffers from accumulating.

We've considered some alternative approaches to this for other operating systems. For one there's posix_fadvise(POSIX_FADV_DONTNEED), but that does more than just writeout dirty data. I've also tried mmap();msync(MS_ASYNC);munmap(); - but at least on linux that doesn't do anything. Using MS_SYNC flushes to disk on linux, but it's synchronous, which isn't what we want here.


I find the sync_file_range() API to be rather useful - so I think it'd make sense to implement it. But baring that, could you possibly clarify somewhere public whether msync(MS_ASYNC) does what we'd need it to do on freebsd? I.e. initiate writeback, without blocking?


Regards,

Andres Freund
Comment 1 Kubilay Kocak freebsd_committer freebsd_triage 2015-10-20 10:57:21 UTC
Thank you for creating this issue report as requested Andres. This issue just got first use of the 'performance' keyword. \o/

Additionally, it appears Hadoop HDFS [1], Redis [2] and MongoDB [3] among others use this system call:

[1] https://issues.apache.org/jira/browse/HDFS-6109
[2] https://github.com/antirez/redis/issues/667
[3] https://jira.mongodb.org/browse/SERVER-18649
Comment 2 crest 2015-10-20 16:37:07 UTC
While not exactly what you're asking for FreeBSD does offer aio_fsync(2) to implement asynchronous file syncing. In theory this offers the kernel the opportunity to flush data to stable storage at the optimal pace and notify the application about it. If we're talking about OS specific code paths its also possible to receive the notification via kqueue(2) + kevent(2) instead of polling or signals. Too bad that the kernel requires that much babysitting from userland.
Comment 3 Warner Losh freebsd_committer 2015-10-20 18:11:22 UTC
making sync_file_range should be straight forward. fsync
is internally implemented as sync_file_range(0, 0, SYNC)
in the kernel right now, so adding the right glue in the kernel
would be easy. well, except for all the hedging of the wording
for the flags...
Comment 4 Warner Losh freebsd_committer 2015-10-20 18:11:42 UTC
making sync_file_range should be straight forward. fsync
is internally implemented as sync_file_range(0, 0, SYNC)
in the kernel right now, so adding the right glue in the kernel
would be easy. well, except for all the hedging of the wording
for the flags...