zmalloc: Out of memory trying to allocate about wrk HOT 14 CLOSED

$ sysctl vm.overcommit_memory
vm.overcommit_memory = 0
$ sysctl vm.overcommit_ratio
vm.overcommit_ratio = 50
$ free
             total       used       free     shared    buffers     cached
Mem:        502944     164716     338228          0      16520      67608
-/+ buffers/cache:      80588     422356
Swap:            0          0          0

Well, I guess it is pretty much self-explanatory... There is simply not enough memory.

from wrk.

@hamiltont can confirm, also on a t1 micro. Thanks!

from wrk.

Hi @suminb, could you share the output of sysctl vm.overcommit_memory and vm.overcommit_ratio on your system? How much RAM and swap does it have?

wrk allocates quite a bit of memory in advance for data samples, though most won't be used during a short test.

from wrk.

It was one of the virtual machines that I use for development purposes. I increased the system memory size to 2 GB and tried it again. It worked without any problem.

from wrk.

Glad to hear you got it running!

from wrk.

For future travelers, this also happens on t1.micro instances. Upgrading to m1.medium resolved the problem as explained above

from wrk.

does anyone know if m1 small is sufficient? I have already set up a small machine and i would rather not to proceed in a medium one... thanks :)

from wrk.

@NickSar I think that wrk allocates most memory up front, so if you can call ./wrk ${URL} (on any URL) without an immediate crash then you should be good to go

from wrk.

any way to reduce the amount of memory allocated by wrk upfront ?

from wrk.

+1 for method to reduce memory up front. For ensuring consistency, perhaps have the flag also print something like WARNING: wrk has been requested to reduce it's memory footprint. You may experience segfaults and/or your results may not be valid

from wrk.

+1

from wrk.

Shorter timeout may save your memory.
https://github.com/wg/wrk/search?utf8=✓&q=stats_alloc&type=Code

from wrk.

ty

from wrk.

39af03f
This commit reduced memory usage dramatically.

Before: 800MB * 2 for statistics.latency and statistics.requests
After: 8 * (timeout [us]) for statistics.latency and 80MB for statistics.requests

from wrk.