Ralf Hildebrandt wrote:
Back in 2009 I was still using 2.7.x: I switched from 2.7.x to 3.1.0.14 yesterday. Today, my proxy would encounter an out of memory condition. cache_mem was EXPLICITLY set to 128 MB, thus the new default of 256 MB would not be relevant. Attached is a picture of squidstats that show the memory consumption before and after the switch. Note the peak on the right edge of the graph. In the end squid was using 1.4GB of RAM. And a retry in 2010: I switched from 3.0.x to 3.1.0.x yesterday. Find some output from % squidclient mgr:mem for 3.0 here http://www.arschkrebs.de/bugs/mgr-mem.3.0.txt.gz and from % squidclient mgr:mem for 3.1 here http://www.arschkrebs.de/bugs/mgr-mem.3.1.txt.gz What I did: 3.0 was running for a whole day, I then used squidclient mgr:mem Then I replaced 3.0 with 3.1 and restarted. I then let it run for about 15 minutes and issued squidclient mgr:mem several times Can anybody see where the excessive memory usage comes from?
3.1 betas had problems with a few things that show the same symptoms as leaks. But were not real leaks, so the memory tools don't detect them well.
* on 64-bit boxes due to non-aligned cache_mem chunk sizes being allocated 2 pages each. Squid only counted the one allocated, so sys shows 2x usage what squid does.
* external_acl_type internal result cache not being capped. Use of the cache=COUNT parameter is needed on high-traffic proxies.
* still open: bug 2305; auth has some logic errors creating circular reference loops. Which could hold onto credentials way too long, or cause real leaks to accumulate under certain request types.
* still open; Squid's custom memory manager (mempools) is "broken" for some reason on 64-bit. Not having looked at the one deeply yet I'm not sure what the impact is.
Can you try your tests on the 3.1.1 (or 3.1.2 in a few days) release please and see if it was one of the solved issues?
Amos -- Please be using Current Stable Squid 2.7.STABLE9 or 3.1.1
