AW: Severe memory leak in global pool: mgm_query
P.S. It seems that this problem doesn't reproduce with 9.40
------------------------------------------
Alexey Sonkin
Senior Database Administrator
> -----Original Message-----
> From: Alexey Sonkin
> Sent: Thursday, August 14, 2003 12:17 PM
> To: 'informix-list@iiug.org'
> Subject: Severe memory leak in global pool: mgm_query
>
> Hi, everybody,
>
> After taking a look at the size of memory pools (onstat -g mem)
> on our machine, I was shocked : it appeared, that the size of global
> pool is 138 MB!!!
>
> Informix Dynamic Server 2000 Version 9.21.UC4 -- On-Line (Prim) -- Up
> 22 days 14:11:09 -- 1790448 Kbytes
>
> Pool Summary:
> name class addr totalsize freesize #allocfrag #freefrag
> .........
> global V 8f816020 134889472 1006776 518347 6502
> .........
>
> This global pool can't be reduced by 'onmode -F'
> (its 'free' size is too small)
>
> I've made an attempt to analyze this pool with
> 'onstat -g afr global'. The resulting file was 16 MB in size
> mostly contained records like:
>
> Allocations for pool name global:
> addr size memid
> .....
> 8f8199d8 240 mgm_query
> 8f819ac8 240 mgm_query
> .....
>
> This is total statistics of that file:
>
> Total_bytes #fragmets fragment_name
> -------------------------------------------
> 40 1 messages
> 48 1 fragman
> 56 1 initseg
> 56 1 shmblklist
> 160 4 Log
> 280 1 filetable
> 280 1 opentable
> 280 1 hashfiletab
> 456 1 mgm
> 568 1 gentcb
> 1496 1 SAPI
> 1832 3 overhead
> 2592 1 ostcb
> 3648 8 rsam
> 4528 11 sql
> 6392 107 misc
> 7432 163 osenv
> 19256 310 mcbmsg
> 41024 2 btcleaner
> 73528 69 scb
> 274944 2 aio
> 559144 139 log
> 3147248 13 datarep
> 5822416 3060 net
> 122795024 509837 mgm_query
> ------------------------------------------
>
> That is, this memory leak is caused by MGM_QUERY fragments:
> We have 509837 fragments 240 bytes each, 122 MB in total !!!!
>
> I have a suspicion, that 'mgm' here stays for 'memory grant manager',
> that is, memory leak is caused by PDQ (parallel data query) queries!
>
> I've attempted to make a search in the PDQ (product defect query) tool,
> and was unable to find anything similar.
>
> By making further experiments with PDQ queries and running simple script
> in parallel (onstat -g afr global | grep mgm_query | wc -l), I've found,
> that each PDQ query allocates from 4 to 1000 mgm_query fragments
> and doesn't release them to the system upon completion!!!
>
> Hey, Informix tech-support, can You file a bug in Your 'Atlas',
> or whatever You use now at IBM?
>
>
> ------------------------------------------
> Alexey Sonkin
> Senior Database Administrator
>
sending to informix-list