Performance degrades with multiple clients connected

[blackberry436]

I was wondering if anyone had seen this issue. We have:
- 3 Cassandra nodes
- using CQL3 query (SELECT * from mytable...)
- mytable has composite column names created with CQL3. i.e.
create table mytable (
userid ascii,
contentid ascii,
read boolean,
location text,
size bigint,
PRIMARY KEY(userid, contentid)
- using hostPoolSize = 10 and using 3 Cassandra hosts
- all async calls in the nodejs server

I am using SoapUI to run tests against the server which is running in AWS. If I use a single thread in the SoapUI load test it averages around 500 ms round trip time to get back our content (about 150k). However, if I use multiple threads in SoapUI we seem to get stuck at downloading 350k bytes per second so the transactions per second keep going down the more threads I add.

We have isolated this to be only an issue when using helenus. When simply downloading static content the same nodejs server can deliver 9 MB per second.

Any comments or help is appreciated...

Thanks.
Carl.

[SimpleSeeker]

I'm also experiencing similar behavior and it seems to be unique to helenus; a coworker tried to do something similar with php and didn't experience the same issues. I alleviate this pain point by caching my results in memcache. Also, from profiling my code with helenus, helenus behaves like it's executing my requests sequentially instead of in parallel.

Duong

[Russell Bradberry]

This is odd and shouldn't be the case.  Helenus isn't doing anything under the covers that would force sequential reading.  This might be a limitation of the Thrift driver though.  As soon as we're done with the native CQL interface, lets test again to see.