WL#643: adaptive block size (malloc optimization)
Affects: Server-7.1 — Status: Assigned
>> After thinking about this for a couple of hours, I belive that in the >> future a good way would be to make the block sizes more adaptive (as >> the malloc would be 'perfect' if we would just know how much memory we >> would really need). A simple way to do this would be something like >> the following: >> >> - Remember for the mem_root how much memory was allocated this time. >> - For each free_root() adjust the block_size as follows: >> >> Adjust 'used_memory' between block_size/2 and block_size*2 >> >> mean_memory_used= (mean_memory_used + used_memory)/2 >> >> block_size= ALIGN_SIZE(mean_memory, 8192); >> >> The above would change block_size to slowly increase / decrease >> according to how the current connection allocates data. It would >> also be quite inexpensive to do. >> >> What do you think about the above ? Sanja> Good idea! Sanja> But I think it is better to count how many memory used, but not +allocated. It would be quite inexpensive to have both and then do some statistics to get the best usage. Sanja> Also very important for us size of preallocated block and it is Sanja> better to adjust it amount (IMHO). Thus memory root can provide interface Sanja> for adjustment, but external procedure will edjust preallocated memory +every Sanja> N calls for example. I.e. use it with calculating block_size, to avoid +very Sanja> frequently allocated second block. Even better to make preallocated block Sanja> adjustments if number of allocation second block will be high then 50% +cases. Sanja> (I mean memory pool of thd here which used to perform command) Sounds good. So please put back malloc optimization to your todo :) Estimation 7hrs (with testing)
Copyright (c) 2000, 2020, Oracle Corporation and/or its affiliates. All rights reserved.