RCU latency measurements

This is again a summary of what was presented at the OLS 2002 RCU paper.

One other aspect of different RCU implementations that we have been investigating is the update latency. That is, how long it takes to complete the grace period and do the actual update. Long latencies could result in system running out of memory. I measured this for 3 different RCU implementations - rcu_poll, rcu_ltimer and rcu_sched against varying number of clients in dbench with the lockfree dcache lookup patch using RCU for dentries.

The results can be seen in the following graph -
http://lse.sourceforge.net/locking/ols2002/rcu/results/latency/latency.png
It is logscale on y axis, in case you don't notice it.

The patches are same as the ones used in overhead measurements -
http://lse.sourceforge.net/locking/ols2002/rcu/patches/

1. rcu_poll, with its forced reschedule and aggressive polling, shows the best latency.

2. The latencies for all these RCU implementations remain reasonably flat under increased load.

Comments/suggestions ?

Thanks

Dipankar Sarma  <[email protected]> http://lse.sourceforge.net
Linux Technology Center, IBM Software Lab, Bangalore, India.