sPHENIX tracking discussion

[Anthony Frawley <afrawley AT fsu.edu>]

Dear Christof and Chris,

TrkrHitTruthAssociation has the same reset instruction. I assume that Christof is running over already produced hits files, so would not see a memory leak from that.

The clear command for a multimap reduces the size to zero. So evidently the problem is that something is not releasing the memory.

I came across a suggestion that may be worth trying:

https://stackoverflow.com/questions/6770316/memory-leak-with-map

The suggestion is to create an empty map (in local scope, presumably) and use swap to exchange the content with persistent map:

The example was:

=============

Try:

{
    std::map<int,int> empty_map;
    empty_map.swap(_map);
}

(At least, this is the usual way to convince a standard library container actually to release its memory.)

=============

It would be worth trying this in the reset method of the container.

Tony

---

From: sPHENIX-tracking-l <sphenix-tracking-l-bounces AT lists.bnl.gov> on behalf of Christof Roland via sPHENIX-tracking-l <sphenix-tracking-l AT lists.bnl.gov>
Sent: Thursday, January 20, 2022 10:17 AM
To: pinkenburg <pinkenburg AT bnl.gov>
Cc: sphenix-tracking-l AT lists.bnl.gov <sphenix-tracking-l AT lists.bnl.gov>
Subject: Re: [Sphenix-tracking-l] TpcClusterizer memory leak (found)

 

Hi Chris, 

ok, I just ran a job explicitely resetting the maps in the clusterizer. This didn't change things. 

See below

Its the Reset() of TrkrClusterHitAssocv3.cc

The filling of it is well contained in an if statement in the clusterizer, so I'll check there if something else is 

alloacted in there.

What is the recommended way of resetting a map?

cheers

   Christof 

On 20. Jan 2022, at 16:00, pinkenburg via sPHENIX-tracking-l <sphenix-tracking-l AT lists.bnl.gov> wrote:

Hi Christof,

the event is cleared in a separate step - after processing by calling the Reset() method for every object under the DST node (we have some opaque foreach type way to iterate over the node tree, what is being executed here is offline/framework/phool/PHNodeReset.cc).

I think the problem is likely in a Reset() method of the container. Which one we are talking about is opaque to me but if I look at the containers in trackbase, they just clear their map without calling the Reset() method of their members. If a member of this map allocates memory it leaks right there.

Chris

On 1/20/2022 7:49 AM, Christof Roland via sPHENIX-tracking-l wrote:

Hi Everybody, 

I tried a few stunts to find the memory leak in the TpcClusterizer and what finally 

stopped it was turning off the creation of the ClusterHitAssoociation map. 

With this disabled the memory along 1000 events is pretty stable. See PrMon plot below.

It looks like thes ClusterHitsAssoc map is not properly reset. Looking in the Clusterizer code 

neither the clustermap nor the assoc map are reset. Which makes sense since the INTT and 

MVTX clusterizers also write to the same maps. 

While ClusterMap apparently gets reset properly, the assoc map doesn't.

Where do we do this actually?

Cheers

   Christof 

<prmon_clusterizer_nohitassoc.jpg>

_______________________________________________
sPHENIX-tracking-l mailing list
sPHENIX-tracking-l AT lists.bnl.gov
https://lists.bnl.gov/mailman/listinfo/sphenix-tracking-l

-- 
*************************************************************

Christopher H. Pinkenburg	;    pinkenburg AT bnl.gov
				;    http://www.phenix.bnl.gov/~pinkenbu

Brookhaven National Laboratory	;    phone: (631) 344-5692
Physics Department Bldg 510 C	;    fax:   (631) 344-3253
Upton, NY 11973-5000

*************************************************************

_______________________________________________
sPHENIX-tracking-l mailing list
sPHENIX-tracking-l AT lists.bnl.gov
https://lists.bnl.gov/mailman/listinfo/sphenix-tracking-l