Mailing-List: contact user-help@zookeeper.apache.org; run by ezmlm
Precedence: bulk
Reply-To: user@zookeeper.apache.org
Received-SPF: neutral (athena.apache.org: local policy)
Content-Type: text/plain; charset=us-ascii
Mime-Version: 1.0 (Apple Message framework v1084)
Subject: Re: Recovery log on /dev/shm?
From: Mahadev Konar <mahadev@hortonworks.com>
In-Reply-To: 
 <CAKR25EQ8oihoVq+o=rw446d0hfXXR17UZv-qFx=tdYsfgEL25A@mail.gmail.com>
Date: Thu, 1 Sep 2011 15:32:49 -0700
Content-Transfer-Encoding: quoted-printable
Message-Id: <349378B8-B08D-4E91-A075-4C98D1C6E0CE@hortonworks.com>
References: 
 <CAF9YjSCrTchTn2gxhqUbkXLG6d6-kcednFEhkTk6ks=pZBbH4A@mail.gmail.com>
 <CAKR25EQ8oihoVq+o=rw446d0hfXXR17UZv-qFx=tdYsfgEL25A@mail.gmail.com>
To: user@zookeeper.apache.org

I had done quite a bit of benchmarks on not logging to disk.
See ZOOKEEPER-866

and noticed that it actually wasnt of much help. It doesnt improve the =
latency by much and also the throughput was still the same.

hope that helps.

mahadev

On Sep 1, 2011, at 2:24 PM, Benjamin Reed wrote:

> have you benchmarked the difference between using the ram disk and
> turning off sync with forceSync=3Dno? i would think that the =
performance
> will be similar. you will need to use the code in trunk because
> https://issues.apache.org/jira/browse/ZOOKEEPER-1087 isn't in any of
> the releases yet.
>=20
> thanx
> ben
>=20
> On Thu, Sep 1, 2011 at 1:31 PM, Ashwin Jayaprakash
> <ashwin.jayaprakash@gmail.com> wrote:
>> This is a follow up to a question I posted in July 2010 - about =
whether the
>> recovery log could be redirected to /dev/null or some such low =
latency drive
>> [http://markmail.org/thread/3edpfmvn2fhyg5wv].
>>=20
>> I just found out that Linux distros with 2.6 kernel have a built-in =
shared
>> memory file system that runs purely in memory - /dev/shm. Since this
>> provides recovery across JVM restarts (but not OS restarts) wouldn't =
this be
>> sufficient to run the ZK nodes off this drive? Replicas are on other
>> machines anyway so FT/HA is already there. It's just that disk =
latency can
>> be eliminated completely.
>>=20
>> [Apologies if you guys are already aware of /dev/shm]
>>=20
>> I ran some simple Java IO tests (not full ZK) on this drive and I =
wanted to
>> share my results. Might be of some use to apps that require low =
latency:
>> =
http://javaforu.blogspot.com/2011/09/ram-disk-is-already-in-linux-and-nobo=
dy.html
>>=20
>> Ashwin.
>>=20