Another piece of this story
I have mocked a dummy LedgerStorage (like DummyLedgerStorage in BK tests)
which does nothing (no real disk storage, non memory, no index), no
checkpoint
The problems happens anyway, after a bunch opf seconds (20-30s) the system
slows down
So the storage is not the p
Any idea?
What happens after the first checkpoint?
It seems that the bookies starts to behave in a different way.
I image that in a first phase all is in memory + journal and then in a
second phase we start using the disk because the amount of data (index +
ledger data) cannot be retained in memory
2017-07-25 15:36 GMT+02:00 Enrico Olivelli :
>
>
> 2017-07-25 13:58 GMT+02:00 Enrico Olivelli :
>
>> I noticed that the "performance" drop in my bench happens contextually to
>> the opening of several "RandomAccessFile", to .dx files.
>>
>> In my bench I continue to perform writes and after some t
2017-07-25 13:58 GMT+02:00 Enrico Olivelli :
> I noticed that the "performance" drop in my bench happens contextually to
> the opening of several "RandomAccessFile", to .dx files.
>
> In my bench I continue to perform writes and after some time the overall
> performance (latency and throughtput) "
I noticed that the "performance" drop in my bench happens contextually to
the opening of several "RandomAccessFile", to .dx files.
In my bench I continue to perform writes and after some time the overall
performance (latency and throughtput) "degrade"
while the bench is running I see that the over
Il lun 24 lug 2017, 19:54 Venkateswara Rao Jujjuri ha
scritto:
> On Mon, Jul 24, 2017 at 3:06 AM, Enrico Olivelli
> wrote:
>
> > 2017-07-21 20:37 GMT+02:00 Enrico Olivelli :
> >
> > >
> > >
> > > Il ven 21 lug 2017, 20:32 Sijie Guo ha scritto:
> > >
> > >> As the discussion in a separate thread
On Mon, Jul 24, 2017 at 3:06 AM, Enrico Olivelli
wrote:
> 2017-07-21 20:37 GMT+02:00 Enrico Olivelli :
>
> >
> >
> > Il ven 21 lug 2017, 20:32 Sijie Guo ha scritto:
> >
> >> As the discussion in a separate thread, it might be making sense to
> check
> >> what is the difference between using pool
2017-07-21 20:37 GMT+02:00 Enrico Olivelli :
>
>
> Il ven 21 lug 2017, 20:32 Sijie Guo ha scritto:
>
>> As the discussion in a separate thread, it might be making sense to check
>> what is the difference between using pooled allocator and unpooled
>> allocator using v3 protocol. Also considering
Il ven 21 lug 2017, 20:32 Sijie Guo ha scritto:
> As the discussion in a separate thread, it might be making sense to check
> what is the difference between using pooled allocator and unpooled
> allocator using v3 protocol. Also considering comparing using heap buffer
> and direct buffer as well.
As the discussion in a separate thread, it might be making sense to check
what is the difference between using pooled allocator and unpooled
allocator using v3 protocol. Also considering comparing using heap buffer
and direct buffer as well.
I am suspecting this might contribute latency.
- Sijie
Kishore, do you have news?
Il ven 14 lug 2017, 09:05 Enrico Olivelli ha scritto:
> At the meeting we told the Kishore will perform some benchmarks on his
> side.
> He will take a look at my code, and we are going to share the results.
> Maybe it will be possible to share the results of benchmark
At the meeting we told the Kishore will perform some benchmarks on his side.
He will take a look at my code, and we are going to share the results.
Maybe it will be possible to share the results of benchmarks done from
Kishore at Salesforce too.
The primary goal is to understand the differences be
2017-07-13 4:11 GMT+02:00 Sijie Guo :
> On Wed, Jul 12, 2017 at 10:35 PM, Enrico Olivelli
> wrote:
>
> > Sijie, JV, just a recap my point of view:
> > - considering latency = "time for asynchAddEntry to complete"
> > - there is a some difference from 4.4 and 4.5 in the usage of memory, but
> > no
On Wed, Jul 12, 2017 at 10:35 PM, Enrico Olivelli
wrote:
> Sijie, JV, just a recap my point of view:
> - considering latency = "time for asynchAddEntry to complete"
> - there is a some difference from 4.4 and 4.5 in the usage of memory, but
> no so clear
> - the type of GC (parallel vs G1) does n
Sijie, JV, just a recap my point of view:
- considering latency = "time for asynchAddEntry to complete"
- there is a some difference from 4.4 and 4.5 in the usage of memory, but
no so clear
- the type of GC (parallel vs G1) does not impact very much but with G1 you
achieve best latency
- using the
On Wed, Jul 12, 2017 at 2:04 AM, Venkateswara Rao Jujjuri wrote:
> Enrico, let me try to paraphrase the issue.
>
> - With G1GC + Netty 4.1 is giving you roughly same perf as prev release. Is
> that accurate statement?
> But you are still seeing latency spikes with Netty 4.1??
> I did not ful
Enrico,
Do you mind share your gc log between your different runs?
- Sijie
On Tue, Jul 11, 2017 at 11:27 PM, Enrico Olivelli
wrote:
> Another interesting thing...during my profiling activity I gave a chance to
> the old v2 protocol and activated the gc logs, as expected the result is
> that wi
Enrico, let me try to paraphrase the issue.
- With G1GC + Netty 4.1 is giving you roughly same perf as prev release. Is
that accurate statement?
But you are still seeing latency spikes with Netty 4.1?? I did not fully
understand your sleep usecase.
How sleep is yielding better latency?
Thanks,
JV
Another interesting thing...during my profiling activity I gave a chance to
the old v2 protocol and activated the gc logs, as expected the result is
that with v2 protocol there is almost no GC activity during the benchmark
-- Enrico
2017-07-11 12:07 GMT+02:00 Enrico Olivelli :
>
>
> 2017-07-11 11
2017-07-11 11:04 GMT+02:00 Sijie Guo :
> I think Netty4 requires more offheap memory. you might need to tune the JVM
> settings. I doubt that latency diff coming from the JVM gc.
>
> A simple thing to verify that is to dump the gc log by adding " -Xloggc:"
> setting and compare the gc logs between
I think Netty4 requires more offheap memory. you might need to tune the JVM
settings. I doubt that latency diff coming from the JVM gc.
A simple thing to verify that is to dump the gc log by adding " -Xloggc:"
setting and compare the gc logs between versions.
- Sijie
On Tue, Jul 11, 2017 at 12:1
a did a bisect and the culprit (in my opinion) is the switch to netty 4 for
the performance regression from 4.5 and 4.4
at commit:
commit 811ece53a1c975c4e768422f3d622ac9de6b3e41BOOKKEEPER-1058: Ignore
already deleted ledger on replication audit
Total time: 204 ms
Total real time: 79 ms per e
Il lun 10 lug 2017, 18:21 Venkateswara Rao Jujjuri ha
scritto:
> With Netty changes, lack of native epoll() has huge perf impact as per
> Kishore.
> Are you sure you are using epoll()?
>
Yes. I tried with netty local transport too. It seems not related to netty
to me.
I will double check, tomorr
With Netty changes, lack of native epoll() has huge perf impact as per
Kishore.
Are you sure you are using epoll()?
On Mon, Jul 10, 2017 at 1:49 AM, Enrico Olivelli
wrote:
> 2017-07-10 10:40 GMT+02:00 Sijie Guo :
>
> > Also one other thing to check is the JVM settings. Do you mind sharing
> that
2017-07-10 10:40 GMT+02:00 Sijie Guo :
> Also one other thing to check is the JVM settings. Do you mind sharing that
> as well?
>
>
this is the surefire config, I am using oracle jdk 8
maven-surefire-plugin
2.20
1
Also one other thing to check is the JVM settings. Do you mind sharing that
as well?
Sijie
On Jul 10, 2017 1:17 AM, "Sijie Guo" wrote:
> I am not sure if there is any default values changed for journal settings.
> I would suggest you testing by setting specifically the journal settings.
>
> Als
2017-07-10 10:17 GMT+02:00 Sijie Guo :
> I am not sure if there is any default values changed for journal settings.
> I would suggest you testing by setting specifically the journal settings.
>
> Also if you can share your benchmark, that would be good for other people
> to verify.
>
Sure
this i
I am not sure if there is any default values changed for journal settings.
I would suggest you testing by setting specifically the journal settings.
Also if you can share your benchmark, that would be good for other people
to verify.
Sijie
On Jul 10, 2017 12:32 AM, "Enrico Olivelli" wrote:
> H
Hi,
I am doing some benchmarks on BK, I see that from 4.4.0 to 4.5.0 there is
something "slow" but I cannot understand what. I really hope that I am
wrong.
I am working with writes, I will pass to reads once writes will be ok.
My problem is both on latency (time for AddComplete callback to complet
29 matches
Mail list logo
