[pve-devel] [RFC PATCH pve-cluster] fix #5728: pmxcfs: allow bigger writes than 4k for fuse

[Dominik Csapak]

On 9/19/24 16:57, Thomas Lamprecht wrote:
> Am 19/09/2024 um 14:45 schrieb Dominik Csapak:
>> On 9/19/24 14:01, Thomas Lamprecht wrote:
>>> Am 19/09/2024 um 11:52 schrieb Dominik Csapak:
>>>> by default libfuse2 limits writes to 4k size, which means that on writes
>>>> bigger than that, we do a whole write cycle for each 4k block that comes
>>>> in. To avoid that, add the option 'big_writes' to allow writes bigger
>>>> than 4k at once.
>>>>
>>>> This should improve pmxcfs performance for situations where we often
>>>> write large files (e.g. big ha status) and maybe reduce writes to disk.
>>>
>>> Should? Something like before/after for benchmark numbers, flamegraphs
>>> would be really good to have, without those it's rather hard to discuss
>>> this, and I'd like to avoid having to do those, or check the inner workings
>>> of the affected fuse userspace/kernel code paths here myself.
>>
>> well I mean the code change is relatively small and the result is rather clear:
> 
> Well sure the code change is just setting an option... But the actual change is
> abstracted away and would benefit from actually looking into..
> 
>> in the current case we have the following calls from pmxcfs (shortened for e-mail)
>> when writing a single 128k block:
>> (dd if=... of=/etc/pve/test bs=128k count=1)
> 
> Better than nothing but still no actual numbers (reduced time, reduced write amp
> in combination with sqlite, ...), some basic analysis over file/write size distribution
> on a single node and (e.g. three node) cluster, ...
> If that's all obvious for you then great, but as already mentioned in the past, I
> want actual data in commit messages for such stuff, and I cannot really see a downside
> of having such numbers.
> 
> Again, as is I'm not really seeing what's to discuss, you send it as RFC after
> all.
> 
>> [...]
>> so a factor of 32 less calls to cfs_fuse_write (including memdb_pwrite)
> 
> That can be huge or not so big at all, i.e. as mentioned above, it would we good to
> measure the impact through some other metrics.
> 
> And FWIW, I used bpftrace to count [0] with an unpatched pmxcfs, there I get
> the 32 calls to cfs_fuse_write only for a new file, overwriting the existing
> file again with the same amount of data (128k) just causes a single call.
> I tried using more data (e.g. from 128k initially to 256k or 512k) and it's
> always the data divided by 128k (even if the first file has a different size)
> 
> We do not override existing files often, but rather write to a new file and
> then rename, but still quite interesting and IMO really showing that just
> because this is 1 +-1 line change it doesn't necessarily have to be trivial
> and obvious in its effects.
> 
> [0]: bpftrace -e 'u:cfs_fuse_write /str(args->path) == "/test"/ {@ = count();} END { print(@) }' -p "$(pidof pmxcfs)"
> 
> 
>>>> If we'd change to libfuse3, this would be a non-issue, since that option
>>>> got removed and is the default there.
>>>
>>> I'd prefer that. At least if done with the future PVE 9.0, as I do not think
>>> it's a good idea in the middle of a stable release cycle.
>>
>> why not this change now, and the rewrite to libfuse3 later? that way we can
>> have some improvements now too...
> 
> Because I want some actual data and reasoning first, even if it's quite likely
> that this improves things Somehow™, I'd like to actually know in what metrics
> and by how much (even if just an upper bound due to the benchmark or some
> measurement being rather artificial).
> 
> I mean you name the big HA status, why not measure that for real? like, probably
> among other things, in terms of bytes hitting the block layer, i.e. the actual
> backing disk from those requests as then we'd know for real if this can reduce
> the write load there, not just that it maybe should.


hi,

first i just wanted to say I'm sorry for my snarky comment about not needing to test
performance for such code. You're right, any insight we can gain there
is good and we (I!) should take the time to do that, even if the
change looks "obvious" like it does here

so i did some benchmarks (mostly disk writes) and wrote the short script below
(maybe we can reuse that?)

----8<----
use strict;
use warnings;

use PVE::Tools;

my $size = shift;

sub get_bytes_written {
         my $fh = IO::File->new("/proc/diskstats", "r");
         die if !$fh;
         my $bytes = undef;
         while (defined(my $line = <$fh>)) {
                 if ($line =~ m/sdb/) {
                         my @fields = split(/\s+/, $line);
                         $bytes = $fields[10] * 512;
                 }
         }
         return $bytes;
}

sub test_write {
         my ($k) = @_;
         system("rm /etc/pve/testfile");
         my $data = "a"x($k*1024);
         system("sync; echo -n 3> /proc/sys/vm/drop_caches");
         my $bytes_before = get_bytes_written();
         PVE::Tools::file_set_contents("/etc/pve/testfile", $data);
         system("sync; echo -n 3> /proc/sys/vm/drop_caches");
         my $bytes_after = get_bytes_written();
         return $bytes_after - $bytes_before;
}

$size //= 128;

my $written = test_write($size) / 1024;
print("$written\n");
---->8----

to simulate our real write patterns with varying file sizes

i installed a fresh pve, turned off pvestatd and put /var/lib/pve-cluster on it's own disk
(/dev/sdb), so diskstats only contains writes from the pmxcfs

the results are below (all sizes are kbytes, ran them multiple times, but they
seem to be consistent)

data size  written (old)  amplification (old)  written (new)  amplification (new)
1          56             56                   56             56
2          72             36                   76             38
4          84             21                   88             22
8          144            18                   104            13
16         236            14                   160            10
32         532            16                   324            10
64         1496           23                   836            13
128        6616           51                   3848           30
256        20600          80                   10568          41
512        87296          170                  43416          84
1024       388460         379                  197032         192

for smaller writes there seems to be a minimum overhead of ~50-100 kbytes
for big files with have a massive amplification of > 100x

my patch does seem to make a difference for files >4k

but the biggest surprise here is that the write amplification
is not linear, but increases with an increase in bytes we want to write
so e.g. going from 128k -> 256k file write we don't just write double the amount,
but 3x to 4x as much.

i also produced a flamegraph according to 
https://www.brendangregg.com/FlameGraphs/cpuflamegraphs.html, but that showed virtually no change 
between versions without
and with my patch (if one has a good svg hoster, i can post them ofc)

so, tl;dr

for small writes does not make that much of a difference, but
we can save ~half the writes for large files in the pmxcfs
(which e.g. a ha-state file could do)