Re: Lilypond Speed

[Peter Chubb]

>>>>> "Nick" == Nick Payne <address@hidden> writes:

Nick> As I have just had a rather powerful evaluation server to play
Nick> around with for a few days while I tested our various Windows
Nick> and Linux server builds on it, I thought I'd also take the
Nick> opportunity to compare the build speed of a reasonably
Nick> substantial score. I used Reinhold's setting of Reubke's sonata
Nick> on the 94th psalm. I tested on three machines, all running the
Nick> same version of Lilypond:

Nick> 1. Dell GX620 workstation, Pentium D dual-core 3.0GHz CPU,
Nick>    1Gb RAM, Ubuntu 9.04 x86: 10min 11sec 
Nick> 2. Dell GX745 workstation, Pentium D dual-core 3.4GHz CPU,
Nick>    2Gb RAM, WinXP SP3: 9min 22sec 
Nick> 3. PowerEdge R710 server, dual quad-core Xeon 5560 2.8GHz CPUs,
Nick>    24Gb RAM, Debian 5 amd64: 4min 4sec


I think you'll find the main difference is in size of L2/L3 cache,
and amount of RAM.  Lily (like many object-oriented programs) tends to
have quite a deep stack, and to use lots of memory --- which it
visits in what looks to the processor like random orders --- so small
caches generate lots of cache misses, which slows things down.  If you
run out of RAM and have to swap, things get even worse.

Xeon 5560: 256k L2, 8M L3 cache (which is almost as fast as the Pentium D's L2 
cache)
Pentium D: 1M L2 cache, no L3 cache.
--
Dr Peter Chubb          www.nicta.com.au    peter DOT chubb AT nicta.com.au
http://www.ertos.nicta.com.au           ERTOS within National ICT Australia