Performance Degradation When Solving Multiple GLOP Problems in Parallel about or-tools

[sschnug]

GLOP is memory intensive. I guess your computer is memory bandwidth bound.

from 64GB it uses only 12Gb for 10 threads, the memory should not be a bottleneck.

That's imho not whats meant by "memory-bound" here.

The question is how much memory needs to be read for doing x cycles of computation.
GLOP being memory intensive could mean, that some operation which is relatively cheap, e.g. an accumulation which would take 1 microsec without waiting for data, but needs to fetch 20MB of memory/data which might be much slower...

Now one can sometimes recognize that something is compute bound (ALU work at cpu is bottleneck) or memory-bound (waiting for memory is bottleneck: e.g. well-optimized matrix-multiplication).

GLOP "seems" to be memory-bound or near that (lots of data is touched for relatively simple calculations) and issues related to that will get much more worse when you overcommit by parallel work as thread-scheduling will lead to cache-invalidation (and maybe less good prefetching) further worsening your memory-fetching.

As this is lots of guessing, one being interested should probably analyze this more by:

• Comparing low-bandwith systems (consumer cpus) with high-bandwith systems (server cpus with many memory-channels)
• Using valgrind or similar "simulation-based" tools
• Using CPU-supported profiling
  • Hardware-counter based analysis of cache-misses and co.

[Mizux]

What is the size of one problem in RAM ? Did you see some swap ?
Also having several GLOP in // you must increase the cache miss ratio between process...

[Diddlik]

What is the size of one problem in RAM ? Did you see some swap ? Also having several GLOP in // you must increase the cache miss ratio between process...

one problem is about 1.1Gb of Ram (regarding TaskManager). I have not observed swaps.
What did you mean by the last sentence?

[lperron]

GLOP is memory intensive.
I guess your computer is memory bandwidth bound.

[Diddlik]

GLOP is memory intensive. I guess your computer is memory bandwidth bound.

from 64GB it uses only 12Gb for 10 threads, the memory should not be a bottleneck.