On Sat, 2020-06-27 at 13:34 -0600, Chris Murphy wrote:
On Sat, Jun 27, 2020 at 11:59 AM Konstantin Kharlamov
> On Sat, 2020-06-27 at 17:00 +0300, Konstantin Kharlamov wrote:
> > Another reason worth mentioning: BTRFS per se is slow. If you look at
> > benchmarks
> > on Phoronix comparing BTRFS with others, BTRFS is rarely even on par with
> > them.
> Btw, I should also add here: it may be clear that in ideal situtation BTRFS
> always be slower than non-COW file systems. The problem however, it is not
> on par with the other open-source COW file system, which is ZFS.
> Some months ago at my dayjob I was performing benchmarks, and out of
> curiosity I
> also compared latest released (as of then, it was 5.6 kernel) BTRFS with
> master of ZFS (which was of a commit b29e31d80 and a kernel 5.4).
> The setup was a RAID5 on 10 SSDs, and a benchmark was three 20-minutes long
> of vdbench with random 70% reads and random 30% writes. For BTRFS I also
> `space_cache=v2` mount option. Results were:
> FS | run 1, IOPS | run 2, IOPS | run 3, IOPS
> BTRFS | 65723.9 | 56474.5 | 55090.2
> ZFS | 96846.1 | 79797.9 | 76249.4
For the sake of the argument I will accept the above as facts.
But raid5 correlates to the desktop how? Does my desktop workload need
96K IOPS? Do I notice the difference even if it can be measured? And
the vdbench command used is? And this particular vdbench command
produces a benchmark that mimics what workload found on the desktop?
There is no hand waving away the relevance of these questions if
you're going to propose performance benchmarks are relevant.
> So, summing up this and my previous mail overall, I do not think
> ordinary desktop BTRFS is currently any good, compared to EXT4 or XFS.
I think your argument is improved if you say we need more and better
benchmarking that mimics the actual workloads we care about; or times
a set of test cases that people can try themselves and reproduce.
I use a variety of OS's on a variety of hardware, including the laptop
I'm using now which is Fedora and Windows dual boot and I'm not
suspicious of performance issues of any kind: Fedora's faster, period.
Might it seem even faster if it were ext4? I've done it, and I don't
think so. So how do you produce a benchmark that accounts for the
user's perception of performance, rather than just raw performance?
Because maybe btrfs is faster. Maybe it's slower. But does it matter?
If it took 5 seconds for GNOME Terminal to launch I'd be mad. 21
seconds is just nonsense. So yes it does matter, but what's the
threshold at which it matters? These benchmarks aren't capturing
either the reality or the feeling. So we need better (more relevant)
benchmarks to have a proper discussion.
For the vast majority of things I'm doing, even if it were the case
that some things are slower, *I* am still much slower than whatever
extra latency there may be. And same goes for the reverse, if btrfs
compression makes some things slightly faster, will I notice? I don't
know. But is that the only metric? I know for sure my hardware is
doing far fewer writes (and reads for that matter), so there is less
wear and tear on the hardware. Is that saving me 50 cents or $50 over
the life of the hardware? I don't know but I do know it's better than
no compression, overall.
Thank you, I see your point. Since it seems to also be present in the reply to
my other mail, let's continue there.