consistency impact on performance

Luke Bakken lbakken at basho.com
Tue Jan 21 14:15:21 EST 2014


Satoshi,

The difference in performance between w=1 and w=3 depends a lot on
your cluster setup and hardware, as well as cluster load.

In your tests, your cluster is most likely under no load, and all
vnodes are ready to process incoming messages. In this case w=1 and
w=3 will have little difference in response time since vnodes will be
able to respond immediately to the write message.

A better test would be to increase cluster load with a tool like
basho_bench and benchmark Riak and Cassandra using that tool.
--
Luke Bakken
CSE
lbakken at basho.com


On Sun, Jan 19, 2014 at 9:55 PM, Satoshi Yamada <bigt2323 at gmail.com> wrote:
> Luke,
>
> Thanks for your reply.
>
> Actually, the size of the data we use for our product is as small as
> hundreds bytes.
> I use such data as 330MB because I thought the difference of performance
> gets clearer by
> storing big data. I did test similarly when the data size is 1MB and 100B,
> but did not
> see much difference either.
>
> thanks,
> Satoshi
>
>
>
>
> 2014/1/20 Luke Bakken <lbakken at basho.com>
>>
>> Hi Satoshi,
>>
>> When using Riak, your object sizes should ideally be 1MB or less. A
>> 330MB object will never result in acceptable Riak performance.
>>
>> If you intend to store large objects like this I strongly recommend
>> using Riak CS, which will break up the object for you into chunks that
>> can be managed by Riak. It also provides an S3-compatible API.
>>
>> Please read more about Riak CS here: http://docs.basho.com/riakcs/latest/
>>
>> Thanks
>> --
>> Luke Bakken
>> CSE
>> lbakken at basho.com
>>
>>
>> On Sun, Jan 19, 2014 at 9:29 PM, Satoshi Yamada <bigt2323 at gmail.com>
>> wrote:
>> > Hi, i'm Satoshi, new to riak.
>> >
>> > I would like to check how consistency changes impacts on the performance
>> > of
>> > riak cluster. I used w=1 and w=3, and I expected at least two times more
>> > of
>> > execution time, but there seems no significant change. I saw more
>> > difference
>> > in Cassandra, so I wonder if it's normal in riak or there is something
>> > wrong
>> > in my testing. Can anyone give me some advice on it?
>> >
>> > I simply checked as shown below.
>> >
>> > w=1
>> > $ time curl -v -XPUT
>> > http://mycluster.com:8098/buckets/w1/keys/data.tar.gz?w=1 -H
>> > "X-Riak-Vclock:
>> > a85hYGB.........." -H "Content-Type: text/plain" --data-binary
>> > @data.tar.gz
>> > ...
>> > ...
>> > ...
>> > real    0m27.501s
>> > user   0m0.378s
>> > sys    0m0.674s
>> >
>> > w=3
>> > $ time curl -v -XPUT
>> > http://mycluster.com:8098/buckets/w3/keys/data.tar.gz?w=3 -H
>> > "X-Riak-Vclock:
>> > a85hYGB.........." -H "Content-Type: text/plain" --data-binary
>> > @data.tar.gz
>> > ...
>> > ...
>> > ...
>> > real    0m29.278s
>> > user   0m0.398s
>> > sys    0m0.674s
>> >
>> > My cluster consists of 40, all active and healthy machines and running
>> > riak-1.4.1.
>> > The data I use is 330MB.
>> >
>> > Thanks in advance,
>> > Satoshi
>> >
>> > _______________________________________________
>> > riak-users mailing list
>> > riak-users at lists.basho.com
>> > http://lists.basho.com/mailman/listinfo/riak-users_lists.basho.com
>> >
>
>




More information about the riak-users mailing list