Running Tornado App on Individual Core

[Kandarp Desai]

Hi,

I've multicore machine to run my tornado app.  After reading few discussions, I've decided to try Supervisor to monitor multiple tornado instances(on same machine) and nginx to balance across those instances. But, I want to assign each instance to individual core, e.g. On 4 cores machine, I want to assign 3 cores to each individual tornado instance running on different ports(8001,8002,8003) and 1 core to other python process. Is it possible to do with supervisor ? Or I have to use some other linux scheduler utilities ? Have anyone tried this kind scenario for their production environment ? Any other recommendation?

Thanks.

Regards,

Kandarp Desai 

[David P. Novakovic]

The operating system will assign the affinity of the processes. There is a fair bit of code in the kernel designed to handle this problem. Especially with regards to the various cache layers on a chip.

I'm assuming you are running a modern processor here :)

David

[David P. Novakovic]

Also further to this, you want to run more than one process per CPU. This will allow greater throughput as there are tiny amounts of time while tornado apps will be waiting for IO from the database.

We have a quad core machine and run 10 instances of tornado, sometimes more.

David

[Kandarp Desai]

Thanks. got it. Basically, I don't need to worry about it :).  Also, I will have to do some benchmarking to do decide exact number of instances required for my app.  

Regards,

Kandarp Desai

[Jeremy Kelley]

Conservative rule-of-thumb (with no scientific backing ymmv iamal
blah) is 1.25-1.5 instances per core you want to "dedicate".

So for an 8 core box, let's say you want to give 7 cores to tornado.
I'd do 7*1.25= 8.75 (round up in this instance) 9 instances.

Again, ymmv.

-j

--
The Christian ideal has not been tried and found wanting;
it has been found difficult and left untried – G. K. Chesterton

[Didip Kerabat]

Pretty much. One thing to note, Nginx proxy_pass doesn't spread the load evenly (It does round robin). If you want to make sure traffic is spread more or less evenly, then you need Nginx fair balancer module:

http://wiki.nginx.org/HttpUpstreamFairModule

https://github.com/gnosek/nginx-upstream-fair

Or use HAProxy

- Didip -

[Phil Plante]

We run 16 instances on a Core i7 2.8ghz quad core with 24gb of ram.  Granted this is just our app server.  We came to this limit because most of the time we're waiting on a Sphinx Search query, or MySql to return data.  Very little time is spent in the Tornado app since its just an API.

[Kandarp Desai]

Thanks Guys for all suggestions :).