In my plain vanilla configs the metrics provider seems to be the
DefaultMetricsProvider
2022-08-28 21:18:59,359 [myid:] - INFO
[main:o.a.z.s.q.QuorumPeerConfig@477] - metricsProvider.className is
org.apache.zookeeper.metrics.impl.DefaultMetricsProvider
I re-ran the 3.8.0 scenario using the NullMetricsProvider and still got
slow results of 16,370 READ tps. Am I misunderstanding your suggestion?
2022-08-28 21:26:48,981 [myid:] - INFO
[main:o.a.z.s.q.QuorumPeerConfig@477] - metricsProvider.className is
org.apache.zookeeper.metrics.impl.NullMetricsProvider
27,052: 3.4.6
15,993:3.6.3
15,943:3.7.1
16,805: 3.8.0, digest.enabled=true:
16,682: 3.8.0: digest.enabled=false
*16,370: 3.8.0 NullMetricsProvider*
On Sun, Aug 28, 2022 at 3:45 AM shrikant kalani <[email protected]>
wrote:
> Hi Will,
>
> I remember there was a discussion in the past that starting 3.6 the
> performance is reduced because of Prometheus metrics endpoint enabled. May
> be you can try disabling metrics.
>
> Recently we compared performance between 3.6.2 and 3.8.0 and found 3.8.0
> performing way better than 3.6.2. There was improvement in both read and
> updates.
>
> Thanks
> Srikant Kalani
>
> On Sun, 28 Aug 2022 at 5:37 PM, tison <[email protected]> wrote:
>
> > Hi Will,
> >
> > Thanks for reporting this case! Could you run the benchmark on 3.7.1
> also?
> > It can help us bisect the point of change.
> >
> > Best,
> > tison.
> >
> >
> > Will Now <[email protected]> 于2022年8月28日周日 05:56写道：
> >
> > > Prior to my server upgrade I'm doing some benchmarking of zookeeper
> > server
> > > 3.4.6 and 3.8.0 using zookeeper-benchmark
> > > https://github.com/brownsys/zookeeper-benchmark
> > >
> > > Methodology: For each server version I extracted it anew from the tar
> > file
> > > onto my linux box, tweaked zoo.cfg to run a single node. On 3.8.0 I
> > enabled
> > > 4 letter words. I run it with a fresh (empty) dataDir. Using default
> > > settings for zookeeper-benchmark. (Contrary to best practices) I am
> > > running the server and test client on the same node :-/
> > >
> > > I perform 3 runs and show the READ average below. I am focussing just
> on
> > > READ results here; results are in reads/sec, so bigger numbers are
> > better:
> > >
> > > 27,052: 3.4.6
> > > 16,805: 3.8.0, digest.enabled=true:
> > > 16,682: 3.8.0: digest.enabled=false
> > >
> > > I saw similar results on a windows box and re-ran on linux for
> additional
> > > confirmation. I find it hard to believe that 3.8.0 is 40% slower on
> > reads!
> > >
> > > A while back I performed similar comparisons of 3.4.6 vs 3.6.x and I
> got
> > > slow results in 3.6.x initially, but disabling digest.enabled fixed it
> > and
> > > the two versions were then comparable. In 3.8.0 I am seeing poor
> results
> > > with or without digest enabled.
> > >
> > > Any thoughts? Are there some new settings I need to tweak? Is my
> > > methodology flawed?
> > >
> > > Thanks!
> > >
> >
>
