Throughput setting in Memaslap (a memcached testing tool) does not seem to work

Question

I am testing the response time of memcached by using memaslap. The memaslap is installed in a lxc (container with 1 core and 1 cache way), and the memcached is installed in another lxc (7 cores with 19 cache ways). I use the default settings for memcached (cachesize = 64M and maxconnection = 1024).

Here is my memaslap settings memaslap -s 192.168.122.123:11211 -S 10s -x 1000000 -T 1 -P 10k.

Then it outputs that Run time: 17.8s Ops: 1000000 TPS: 56306 Net_rate: 57.2M/s.

However, when I use another throughput setting:memaslap -s 192.168.122.123:11211 -S 10s -x 1000000 -T 1 -P 100k.

Then it reprints the almost same result Run time: 17.9s Ops: 1000000 TPS: 55760 Net_rate: 56.5M/s.

Does anyone can tell the reason why throughtput setting does not work? My testing platfrom is Intel(R) Xeon(R) CPU E5-2609 v4 @ 1.70GHz， one cpu socket with 8 physical cores, not supporting hyper-thread, and 16G memory.

Answer 1

Using -P|--tps sets "expected TPS" per connection, while number of connections is set with -c|--concurrency.

Each thread (number of threads set with -T|--threads) will open concurrency/threads connections and watch each connections throughput, tps/concurrency.

Now, when performing the load test, and the actual throughput exceeds the expected throughput, the current thread will sleep for 10µs.

That's all.