Throughput issue contd. On Joyend Riak Smartmachine

Reid Draper reiddraper at gmail.com
Wed Jun 27 07:55:24 EDT 2012


On Jun 27, 2012, at 7:48 AM, Yousuf Fauzan wrote:

> This is great.
> 
> I was loading data using Python. My code would spawn 10 threads and put data in a queue. All threads would read data from this queue.
> However, all threads were hitting the same server/load balancer.
> 
> I tried a different setup too. Where I spawned processes with each process having its own queue. In this case too, all processes were hitting the same server.
> 
> I just now made a change to my code. So now I have 10 threads randomly selecting a node and storing data in it.
> Again, I am getting around 50 writes/sec

When the threads randomly pick a node, do they create a new connection to it, or do they pull the connection from
a pool? As you saw with the throughput difference between curl and python, persistent connections make
big difference.

> 
> Could there be something wrong with the way I have written my loader script?
> 
> On Wed, Jun 27, 2012 at 5:10 PM, Russell Brown <russell.brown at mac.com> wrote:
> 
> On 27 Jun 2012, at 12:36, Yousuf Fauzan wrote:
> 
>> So I changed concurrency to 10 and put all the IPs of the nodes in basho bench config.
>> Throughput is now around 1500.
>> 
> 
> I guess you can now try 5 or 15 concurrent workers and see which is optimal for that set up to get a good feel for the sizing of any connection pools for your application.
> 
> You can also see how adding nodes and adding workers effects your results to help you size the cluster you need for your expected usage.
> 
> Cheers
> 
> Russell
> 
>> 
>> On Wed, Jun 27, 2012 at 4:40 PM, Russell Brown <russell.brown at mac.com> wrote:
>> 
>> On 27 Jun 2012, at 12:09, Yousuf Fauzan wrote:
>> 
>>> I used examples/riakc_pb.config
>>> 
>>> {mode, max}.
>>> 
>>> {duration, 10}.
>>> 
>>> {concurrent, 1}.
>> 
>> Try upping this. On my local 3 node cluster with 8gb ram and an old, cheap quad core per box I'd set concurrency to 10 workers.
>> 
>>> 
>>> {driver, basho_bench_driver_riakc_pb}.
>>> 
>>> {key_generator, {int_to_bin, {uniform_int, 10000}}}.
>>> 
>>> {value_generator, {fixed_bin, 10000}}.
>>> 
>>> {riakc_pb_ips, [{<IP of one of the nodes>}]}.
>> 
>> I add all the IPs here, one entry per node.
>> 
>>> 
>>> {riakc_pb_replies, 1}.
>>> 
>>> {operations, [{get, 1}, {update, 1}]}.
>>> 
>>> 
>>> On Wed, Jun 27, 2012 at 4:37 PM, Russell Brown <russell.brown at mac.com> wrote:
>>> 
>>> On 27 Jun 2012, at 12:05, Yousuf Fauzan wrote:
>>> 
>>>> I did use basho bench on my clusters. It should throughput of around 150
>>> 
>>> Could you share the config you used, please?
>>> 
>>>> 
>>>> On Wed, Jun 27, 2012 at 4:24 PM, Russell Brown <russell.brown at mac.com> wrote:
>>>> 
>>>> On 27 Jun 2012, at 11:50, Yousuf Fauzan wrote:
>>>> 
>>>>> Its not about the difference in throughput in the two approaches I took. Rather, the issue is that even 200 writes/sec is a bit on the lower side.
>>>>> I could be doing something wrong with the configuration because people are reporting throughputs of 2-3k ops/sec
>>>>> 
>>>>> If anyone here could guide me in setting up a cluster which would give such kind of throughput.
>>>> 
>>>> To get the kind of throughput I use multiple threads / workers. Have you looked at basho_bench[1], it is a simple, reliable tool to benchmark Riak clusters?
>>>> 
>>>> Cheers
>>>> 
>>>> Russell
>>>> 
>>>> [1] Basho Bench - https://github.com/basho/basho_bench and http://wiki.basho.com/Benchmarking.html
>>>> 
>>>>> 
>>>>> Thanks,
>>>>> Yousuf
>>>>> 
>>>>> On Wed, Jun 27, 2012 at 4:02 PM, Eric Anderson <anderson at copperegg.com> wrote:
>>>>> On Jun 27, 2012, at 5:13 AM, Yousuf Fauzan <yousuffauzan at gmail.com> wrote:
>>>>> 
>>>>>> Hi,
>>>>>> 
>>>>>> I setup a 3 machine riak SM cluster. Each machine used 4GB Ram and riak OpenSource SmartMachine Image.
>>>>>> 
>>>>>> Afterwards I tried loading data by following two methods
>>>>>> 1. Bash script
>>>>>> #!/bin/bash
>>>>>> echo $(date)
>>>>>> for (( c=1; c<=1000; c++ ))
>>>>>> do
>>>>>> 	curl -s -d 'this is a test' -H "Content-Type: text/plain" http://127.0.0.1:8098/buckets/test/keys
>>>>>> done
>>>>>> echo $(date)
>>>>>> 
>>>>>> 2. Python Riak Client
>>>>>> c=riak.RiakClient("10.112.2.185") 
>>>>>> b=c.bucket("test")
>>>>>> for i in xrange(10000):o=b.new(str(i), str(i)).store()
>>>>>> 
>>>>>> For case 1, throughput was 25 writes/sec
>>>>>> For case 2, throughput was 200 writes/sec
>>>>>> 
>>>>>> Maybe I am making a fundamental mistake somewhere. I tried the above two scripts on EC2 clusters too and still got the same performance.
>>>>>> 
>>>>>> Please, someone help
>>>>> 
>>>>> 
>>>>> The major difference between these two is the first is executing a binary, which has to basically create everything (connection, payload, etc) every time through the loop.  The second does not - it creates the client once, then iterates over it keeping the same client and presumably the same connection as well.  That makes a huge difference.
>>>>> 
>>>>> I would not use curl to do performance testing.  What you probably want is something like your python script that will work on many threads/processes at once (or fire them up many times).
>>>>> 
>>>>> 
>>>>> Eric Anderson
>>>>> Co-Founder
>>>>> CopperEgg
>>>>> 
>>>>> 
>>>>> 
>>>>> 
>>>>> 
>>>>> _______________________________________________
>>>>> riak-users mailing list
>>>>> riak-users at lists.basho.com
>>>>> http://lists.basho.com/mailman/listinfo/riak-users_lists.basho.com
>>>> 
>>>> 
>>> 
>>> 
>> 
>> 
> 
> 
> _______________________________________________
> riak-users mailing list
> riak-users at lists.basho.com
> http://lists.basho.com/mailman/listinfo/riak-users_lists.basho.com

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.basho.com/pipermail/riak-users_lists.basho.com/attachments/20120627/3ec39bc9/attachment.html>


More information about the riak-users mailing list