[
https://issues.apache.org/jira/browse/TS-2796?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13996664#comment-13996664
]
Alan M. Carroll commented on TS-2796:
-------------------------------------
We know that the ram cache was significantly modified for 4.0 so it's plausible
an issue was introduced. A key point here is that the IOBufferBlocks in the ram
cache originally came from the CacheVC. When a CacheVC needs to read a fragment
from disk, it gets an IOBufferBlock in the line Brian noted and sets the I/O to
read to that memory. If at a later point it decides to put the fragment in the
ram cache, it simply moves the IOBufferBlock there, it doesn't copy or re-read
from disk. Therefore if the ram cache is not releasing the IOBufferBlocks as
expected (or they are getting dropped somewhere in the transfer process) this
is what we would see - growing memory usage from putatively in use
IOBufferBlocks. In that case reclaimable free lists won't help.
> Leaking CacheVConnections
> -------------------------
>
> Key: TS-2796
> URL: https://issues.apache.org/jira/browse/TS-2796
> Project: Traffic Server
> Issue Type: Bug
> Components: Cache
> Affects Versions: 4.0.2, 4.2.1, 5.0.0
> Reporter: Brian Geffon
> Labels: yahoo
> Fix For: 5.0.0
>
>
> It appears there is a memory leak in 4.0.x, 4.2.x, and master leaking
> CacheVConnections resulting in IOBufAllocator leaking also, here is an
> example:
> allocated | in-use | type size | free list name
> 67108864 | 0 | 2097152 |
> memory/ioBufAllocator[14]
> 67108864 | 19922944 | 1048576 |
> memory/ioBufAllocator[13]
> 4798283776 | 14155776 | 524288 |
> memory/ioBufAllocator[12]
> 7281311744 | 98304000 | 262144 |
> memory/ioBufAllocator[11]
> 1115684864 | 148242432 | 131072 |
> memory/ioBufAllocator[10]
> 4974444544 | 379977728 | 65536 |
> memory/ioBufAllocator[9]
> 9902751744 | 5223546880 | 32768 |
> memory/ioBufAllocator[8]
> 14762901504 | 14762311680 | 16384 |
> memory/ioBufAllocator[7]
> 6558056448 | 6557859840 | 8192 |
> memory/ioBufAllocator[6]
> 41418752 | 30502912 | 4096 |
> memory/ioBufAllocator[5]
> 524288 | 0 | 2048 |
> memory/ioBufAllocator[4]
> 0 | 0 | 1024 |
> memory/ioBufAllocator[3]
> 0 | 0 | 512 |
> memory/ioBufAllocator[2]
> 32768 | 0 | 256 |
> memory/ioBufAllocator[1]
> 0 | 0 | 128 |
> memory/ioBufAllocator[0]
> 2138112 | 2124192 | 928 |
> memory/cacheVConnection
> [~bcall] has observed this issue on 4.0.x, and we have observed this on 4.2.x.
> The code path in CacheVC that is allocating the IoBuffers is
> memory/IOBuffer/Cache.cc:2603; however, that's just the observable symptom
> the real issue here is the leaking CacheVC.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
