Bug 61087

Our organization occasionally runs fairly large-scale load tests involving 10+ load generation blades in multiple blade centers against large infrastructures.  During one particular large scale load test involving SAML flows, the bottleneck in the load testing suite was ultimately the network bandwidth to the blade centers housing the load generator "master" node that was receiving all the simple data writer (SDW) CSV data.  Upon investigation it was determined that even though we were using "Stripped Async" mode for the SDW, which strips off the RESPONSE data, it was still sending all the REQUEST data (including a somewhat large SAML document) back to the master. This required us to run multiple masters each with 3 slaves in order not to eat up a Gigabit ethernet link to the masters.  However, the blade centers 2 Gigabit uplinks then ultimately became saturated.

It is my belief that the SDW should only be sending the actual CSV data to the master.  If my understanding is wrong, perhaps I ought to look into creating my own custom "CSV Data Writer" or some such thing...

We are planning to capture some tcp dumps to verify this on the latest version of JMeter and will post more information when we get it.

Thank you,
Tim