Re: ext4: performance regression introduced by the cgroup writeback support

[Chris Mason <clm@fb.com>]

On Wed, Sep 23, 2015 at 01:49:31PM +0000, Dexuan Cui wrote:
> Hi all,
> Since some point between July and Sep, I have been suffered from a strange "very slow write" issue and on Sep 9 I reported it to LKML (but got no reply): https://lkml.org/lkml/2015/9/9/290
> 
> The issue is: under high CPU and disk I/O pressure, *some* processes can suffer from a very slow write speed (e.g., <1MB/s or even only 20KB/s), while the normal write speed should be at least dozens of MB/s.
> 
> I think I identified the commit which introduced the regression:
> ext4: implement cgroup writeback support (https://git.kernel.org/cgit/linux/kernel/git/next/linux-next.git/commit/?id=001e4a8775f6e8ad52a89e0072f09aee47d5d252)
> 
> This commit is already in the mainline tree, so I can reproduce the issue there too:
> With the latest mainline,  I can reproduce the issue; after I revert the patch, I can't reproduce the issue.
> 
> When the issue happens:
> 1. the read speed is pretty normal, e.g.. it's still >100MB/s.
> 2. 'top' shows both the 'user' and 'sys' utilization is about 0%, but the IO-wait is always about 100%.
> 3. 'iotop' shows the read speed is 0 (this is correct because there is indeed no read request)  and the write speed is pretty slow (the average is <1MB/s or even 20KB/s).
> 4. when the issue happens, sometimes any new process suffers from the slow write issue, but sometimes it looks not all the new processes suffers from the issue.
> 5. The " WARNING: CPU: 7 PID: 6782 at fs/inode.c:390 ihold+0x30/0x40() " in my Sep-9 mail may be another different issue.
> 6. To reproduce the issue, I need to run my workload for enough long time (see the below).
> 
> My workload is simple: I just repeatedly build the kernel source ("make clean; make -j16"). My kernel config is attached FYI.
> 
> I can reproduce the issue on a physical machine: e.g., in my kernel building test with my .config, it took only ~5 minutes in the first 176 runs, but since the 177th run, it could take from 10 hours to 5 minutes - very unstable.
> 
> It looks it's easier to reproduce the issue in a Hyper-V VM: usually I can reproduce the issue within the first 10 or 20 runs.
> 
> Any idea?

Are you using cgroups?  That patch really shouldn't impact load unless
there are actual IO controls in place.

-chris