Exploring Riak, need to confirm throughput

Matthew MacClary macclary at lifetime.oregonstate.edu
Thu Apr 4 16:14:12 EDT 2013


Thanks for the feedback. I made two changes to my test setup and saw better
throughput:

1) Don't write to the same key over and over. Updating a key appears to be
a lot slower than creating a new key

2) I used parallel PUTs

The throughput I was measuring before was about 26MB/s on localhost. With
these changes it went to around 200MB/s on a disk that can write at about
480MB/s. That is more the type of performance I need for the data store we
have in mind. I am going to proceed with testing on 8 nodes with RAID0
drives.

Here are some details of the testing I did if it will help others. I tried
the test with 1MB, 10MB, and 20MB binary data. I didn't notice a big signal
with regard to larger objects slowing things down.

wget
http://downloads.basho.com.s3-website-us-east-1.amazonaws.com/riak/1.2/1.2.1/rhel/5/riak-1.2.1-1.el5.x86_64.rpm

sudo rpm -Uvh riak-1.2.1-1.el5.x86_64.rpm
/usr/sbin/riak start
mkdir data-dir && cd data-dir
seq -w 0 100 | parallel dd if=/dev/zero of={}.10meg bs=8k count=1280
http_proxy=   # don’t contact proxy
time find . -name \*.10meg | parallel -j8 -n1 wget --post-file {}
http://127.0.0.1:8098/riak/test1/{}

During these tests I saw beam.smp jumping to 350-550 while watching %CPU
under top. When I was seeing slower thoughput beam.smp was using much less
CPU.

Kind regards,

-Matt

On Wed, Apr 3, 2013 at 7:20 AM, Reid Draper <reiddraper at gmail.com> wrote:

> inline:
>
>
> On Apr 2, 2013, at 6:48 PM, Matthew MacClary <
> macclary at lifetime.oregonstate.edu> wrote:
>
> Hi all, I am new to this list. Thanks for taking the time to read my
> questions! I just want to know if the data throughput I am seeing is
> expected for the bitcask backend or if it is too low.
>
> I am doing the preliminary feasibility study to decide if we should
> implement a Riak data store. Our application involves rendering chunks of
> data that range in size from about 1MB-9MB or so. This rendering work is
> CPU intensive so it is spread over a bunch of compute nodes which write the
> output into a data store.
>
>
> Riak is not intended to store objects of this size, not at the moment
> anyway. Riak CS [1], on the other hand, can store files up to several TB.
> That being said, Riak CS may or may not have other qualities  you desire.
> It's a known issue [2] that the Riak object size limitations should be
> better documented.
>
>
> After rendering, a second process consumes that data chunks from the data
> store at a rate of about 480MB/s in a streaming configuration so there is >
> 480MB/s of new data coming in at the same time the data is being read.
>
>
> Is this a single-socket, or is there some concurrency here?
>
>
> My testing so far involves a one node cluster on a dev box. What I wanted
> to show is that Riak writes were limited by the hard disk throughput. So
> far I haven't seen writes to localhost come anywhere close to the hard disk
> throughput:
>
> $ MYFILE=/tmp/output.png
> $ dd if=/dev/zero of=$MYFILE bs=8k count=256k
> 262144+0 records in
> 262144+0 records out
> 2147483648 bytes (2.1 GB) copied, 4.48906 seconds, 478 MB/s
> $ rm $MYFILE
>
> So the hard disk throughput is around 478MB/s for this simple write test.
>
> The next test I did was to load a 39MB binary file into my one node
> cluster. I used a script to do 12 POSTs with curl and 12 POSTSs with wget.
>
> curl --tcp-nodelay -XPOST http://${IP}:${PORT}/riak/test/file3 \
>     -H "Content-Type:application/octet-stream" \
>     --data-binary @${UPLOAD_FILE} \
>     --write-out "%{speed_upload}\n"
>
> wget --post-file ${UPLOAD_FILE} http://127.0.0.1:8098/riak/test/file1
>
> What I found was that I could get only about 26MB/s with this command line
> testing. Does this seam about right? Should I see an 18x slow down over the
> write speed of the disk drive?
>
>
> Was this running the 24 (12 * 2) uploads in serial or parallel? With a
> single-threaded workload, you're unlikely to get Riak to be able to
> saturate a disk. Furthermore, there are design decisions in Riak at the
> moment that make it less than optimal for single objects of 39MB.
> Single-object high throughput (measured in MB) is more in the wheelhouse of
> Riak CS than Riak on it's own, which is primarily designed for low-latency
> and high-throughput (measured in ops/sec). One of the ways that Riak CS
> achieves this on top of Riak is by introducing concurrency between the
> end-user and Riak.
>
>
> Thanks for your comments on my application and test approach!
>
>
> Hope this helps,
> Reid
>
> [1] http://docs.basho.com/riakcs/latest/
> [2] https://github.com/basho/basho_docs/issues/256
>
>
>
> -Matt
>
> -----------------------------------------------
> Dev Environment Details:
> dev box  running RHEL6.2, 12 cores, 48GB, 6Gb/s SAS 15k HD
> Riak 1.2.1 from
> http://downloads.basho.com.s3-website-us-east-1.amazonaws.com/riak/1.2/1.2.1/rhel/5/riak-1.2.1-1.el5.x86_64.rpm
> n_val=1
> r=1
> w=1
> backend=bitcask
>
> Deploy Environment Details:
>  Node to node bandwidth > 40Gb/s
>  similar config for node servers
>  n_val=3
>  r=1
>  w=1
>  backend=?
> _______________________________________________
> 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/20130404/38b4520d/attachment.html>


More information about the riak-users mailing list