TLB impact on VM performance

[Aleksey Shipilev]

On 11/28/2012 10:14 AM, Xin Tong wrote:
> I recently took some data on the performance of hotspot when hugepage
> is enabled/disabled. i found a performance difference of ~20% on
> specjbb with -Xms1024m -Xmx2048m running 4 whs. the nursery heap size
> is ~600M and tenure is ~500M.
> 
> OS: Linux socrates 3.2.0-33-generic #52-Ubuntu SMP Thu Oct 18 16:29:15
> UTC 2012 x86_64 x86_64 x86_64 GNU/Linux
> CPU: model name      : Intel(R) Core(TM) i5-2410M CPU @ 2.30GHz
> 
> I am wondering does hugepage demostrate performance improvements on
> other benchmarks as well ? i see a much smaller difference for dacapo
> benchmarks.

My interpretation: SPECjbb is known to be very memory-access intensive,
that is why improving TLB accesses gives you huge payoffs. Many Dacapo
benchmarks are compute-bound, and the TLB performance while important is
not having the same impacts.

Many characteristic Java applications are as promiscuous with memory as
SPECjbb, and so large pages demonstrate significant improvements on
business workloads we have internally.

-Aleksey.