Memory Pool performance

[Karlos]

Hi all,

Does anybody know if there is a golden puddlesize / threshold size ratio for exec v39+ pooled memory allocator?

I am fine tuning some low level code that needs to allocate/free lots of small objects as well has handling less frequent large allocations. Naturally the memory pool stands out as the OS provided mechanism for doing this without fragmenting the ram to bits.

I understand the CreatePool() parameters. My question is, are there any performance (read speed) concerns relating to the puddle size v threshold size?

As for choosing sensible values, my initial thought was to use the system's page size for the puddle size, but on reflection there is no guarentee that the puddles will be page aligned in any case.

Any suggestions?

[itix]

Well...

Allocating new puddles can be time consuming. Thus if you are doing lot of allocations try minimize need of new puddles.

And for puddle sizes... I think it is better choose something like 4000 than 4200. Puddle sizes are always rounded up (if at all, on OS3 they arent).

[Karlos]

Ok, makes sense.

When you allocate a large amount of memory (ie many times larger than the puddlesize), does that allocation take much longer than the equivalent AllocVec() call?

My understanding of the system is that it just allocates memory rounded to the puddlesize and allows sub allocations for several small objects within a single puddle..

-edit-

As for sizes, I was thinking to use a power of 2, eg 4096 or 8192, but that assumes the system doesnt store any info in the puddle itself.

The only documentation I have for the pool system is in the 3.5 includes&autodocs. Are there any better references available?

[itix]

In general AllocPooled() is not probably any slower than AllocMem(). But of course if Exec finds out that all puddles are full then Exec must call AllocMem() as well. Thus it is always better oversize puddles bit rather than undersize.

Another thing is FreePooled(). FreeMem() is very simple call but FreePooled() must go through all puddles. In the worst case scenario calling FreePooled() can be very demanding. Piru investigated this issue.. Must ask him :-D

My understanding of the system is that it just allocates memory rounded to the puddlesize and allows sub allocations for several small objects within a single puddle..

Correct. If memory allocation is larger than treshold value then new puddle is allocated.