CoCoA » CoCoALib

I am reasonably sure that the upper limit is far too large: as it currently stands it can ask the system for 64M of contiguous RAM.

Probably the lower limit is needlessly large too: even 1K or 4K may be enough. Note that Singular's memory manager (called omalloc, I believe) works with 4K pages.

Note that if we do want to allow a MemPool to free some of its loaves then it is easier for a small loaf to be freeable than for a large one -- related to #437

The problem arose when I experimented creating all (prime) finite fields up to characteristic 32767. There are about 3500 such fields. Finding all primes took less than 0.01s, but creating all the fields took about 0.35s. The time to create the fields was primarily due to the creation of one MemPool for each field. Also the space occupied was about 200Mbytes.

Reducing the smallest MemPool loaf to 4Kbytes, reduced the creation time to about 0.05s.

Independently of the above, in the end, I have decided to discard the idea of automatically creating all small prime finite fields at start up -- see #785.

Someone should do some benchmarking to see if changing the lower/upper limits for loaf size has any measurable effect on run-time performance.
[JAA guesses that there will be almost no measurable effect]

I think that implicit could be a good test.
I can do it, could you give me indications on the first tests to try? (and which lines I should change?)

Thanks, Anna, for volunteering!

The relevant lines are in MemPool.C around line 100.

Change MinLoafBytes to 4*1024. Maybe you could also try 1024.
Change MaxLoafBytes to 64*1024. Maybe you could also try 1024*1024.

That's FOUR sets of experiments to try. My guess is that they will all
come out roughly the same for computation time.

about to try..
(some compilation problems...)

Are the problems my fault? Or some problem with the interface of MemPool? :-(

John Abbott wrote:

Are the problems my fault? Or some problem with the interface of MemPool? :-(

no, my problems. And then I was busy. Doing the tests again right now. Conclusions soon

John Abbott wrote:

Change MinLoafBytes to 4*1024. Maybe you could also try 1024.
Change MaxLoafBytes to 64*1024. Maybe you could also try 1024*1024.

I made several combinations on an example taking approx 2 mins and creating 950 rings.
I cannot see any difference in the timings for 1, 4, 64...

John Abbott wrote:

Did you also try tests which create lots of ring elements? Maybe a non-trivial GBasis?

I tried test-implicit2: now I see I tried only the "direct" algorithm.
Now I try with elimth as well.