Hello!
On Wed, Feb 26, 2014 at 2:41 AM, snarapureddy wrote:
We are using nginx for file uploads instead of directing to the backend
servrs. Used lua openresty module to get the data in chunks in write it to
local disk. File size could vary from few KB’s to 10MB.
Ensure you’re interleaving disk writes and network reads. The
recommended way is to read a chunk from the network and write it
immediately to the file system and again and again.
We are tuning worker process, connections, accept_mutex off etc, but if we
cuncerrently upload files some of the connections were very slow.
Chuck size is 4096.
You can try a bit larger chunk sizes like 16KB or 32KB.
CPU utilization is very minimal. We are running 10
worker processes, but most of the cases sam process is handling multiple
connections and they are becoming slow.
This sounds like a good analysis candidate for the off-CPU flame graph
tool:
https://github.com/agentzh/nginx-systemtap-toolkit#sample-bt-off-cpu
This can help us identify exactly what the bottleneck is in your nginx.
Also, measuring the epoll loop blocking latency distribution could be
insightful too:
https://github.com/agentzh/stapxx#epoll-loop-blocking-distr
Try running more nginx worker processes if file IO syscalls are
blocking your workers too much.
BTW, you may get more and faster responses if you post such questions
on the openresty-en mailing list:
https://groups.google.com/group/openresty-en
Best regards,
-agentzh