zookeeper-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Alexander Shraer <shra...@gmail.com>
Subject Re: Write throughput is low !!
Date Thu, 14 Aug 2014 17:26:15 GMT
Hi,

I'd expect the throughput with 5 servers to be around 10K ops/sec for 100%
writes.

The graph Jordan sent shows smaller numbers than in the ZK paper (Figure 6:
https://www.usenix.org/legacy/event/usenix10/tech/full_papers/Hunt.pdf) or
what I saw with 7 servers here (Figure 6):
https://www.usenix.org/system/files/conference/atc12/atc12-final74.pdf

These were generated using our system test (see src/java/systest/README.txt
on how to run).
If you do run it, I suggest that you use the version in trunk, not 3.5.0
due to ZK-2008 that is fixed there.


Alex



On Thu, Aug 14, 2014 at 9:09 AM, Jordan Zimmerman <
jordan@jordanzimmerman.com> wrote:

> Here is the standard graph of ZK tps:
>
> http://zookeeper.apache.org/doc/r3.1.2/images/zkperfRW.jpg
>
> If you are doing 100% writes with 5 ZK instances, 5K tps sounds right. You
> can get more tps if you switch to 3 instances. The real question, though,
> is why are you doing so many writes?
>
> -Jordan
>
>
> From: Mudit Verma <mudit.f2004912@gmail.com>
> Reply: user@zookeeper.apache.org <user@zookeeper.apache.org>>
> Date: August 14, 2014 at 10:50:53 AM
> To: user@zookeeper.apache.org <user@zookeeper.apache.org>>
> Subject:  Write throughput is low !!
>
> Hi All,
>
> We, at our company, have a modest Zookeeper lab setup with 5 servers.
>
> During our lab tests, we found that somehow we can not go beyond 5K
> Writes/Sec in total with roughly 100 clients (all the clients issues writes
> to different znodes in a tight loop). Can someone point me to the official
> numbers? Is this a good number for Zookeeper or it can scale much beyond?
>
> I understand that, parallelization by writing to different znodes will not
> increase the throughput as there is only one log underneath.
>
> PS: Lab is hosted on a high speed network (1gigbit/sec) and network
> latencies are far lower in contract to the the numbers we have.
>
>
> Thanks
> Mudit
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message