[erlang-questions] Hidden Memory Hog

[Mihai Balea]

On Jul 31, 2011, at 7:31 PM, James Aimonetti wrote:

> -----BEGIN PGP SIGNED MESSAGE-----
> Hash: SHA1
> 
> List,
> 
> I'm at my wits end (they're short, no doubt, but still). I have a VM
> running with several OTP applications that we've written. We see, over
> the course of the day, memory consumption go up until its taking ~30%
> system memory and 300-500MB of swap. Iterating over all processes and
> checking their total_heap_size, no single process was greater than ~1
> MB, with ~300 processes runnning. Forcing garbage collection on all
> processes released maybe 15MB of system memory.
> 
> erlang:memory() lists total memory around 18-20MB; ets tables were all
> minimal, no dets tables. No processes has a message_queue_len > 0. No
> processes that we maintain use the process dictionary.
> 
> I started a second VM, migrating our OTP apps to the second VM, one at a
> time, hoping that stopping one would cause a release of lots of memory
> on the first VM. No such luck; it released a couple more MBs but still a
> large presence and no change in swap usage. I then stopped everything I
> could until only stdlib and kernel were running on the first VM, but to
> no avail.
> 
> Each application has 1-3 long running gen_servers that listen on an AMQP
> queue and spawn worker processes to handle messages. They carry almost
> no state. Some of the handlers will read binary blobs, open a socket,
> wait for a connection, then transmit the blobs, closing the socket and
> dying afterwards (they timeout after 5 minutes if no connection is
> attempted, and are in a supervision tree so I know there aren't a mass
> of them lying around). The blobs are between 20KB and 2MB. The rest of
> the workers are similarly ephemeral.
> 
> What other options are there to find where memory has been allocated? We
> use binaries almost exclusively for strings, don't construct large
> lists. I've poured over the efficiency guide looking for ideas of where
> we've gone astray but am coming up blank so far.
> 
> The system is idle; the only processes should be (and are, as far as I
> can tell) the gen_servers meant to run for a long time.
> 
> Erlang R14B01 (erts-5.8.2) [source] [64-bit] [smp:4:4] [rq:4]
> [async-threads:8] [kernel-poll:true]
> 
> Any ideas are welcome,

Check the output of erlang:memory/0
Binaries, ets tables, atoms all use separate heaps that are different than your regular process heaps.
Look for any of those that appears particularly large.

Mihai