[tor-relays] Tor Relay on ARM server Marvell Armada 370/XP

[Roman Mamedov]

On Tue, 20 Dec 2016 14:24:47 +0100
"Fabio Pietrosanti (naif) - lists" <lists at infosecurity.ch> wrote:

> Hello,
> 
> i'm experimenting Tor setup on very cheap servers from scaleaway.com
> that run a quad-core ARM Marvell Armada 370/XP servers but have
> unlimited bandwidth.
> 
> Those are Soc platform:
> http://natisbad.org/NAS2/refs/Marvell_ARMADA_370_SoC.pdf
> 
> I saw no information on torproject mailing list about those SOC.
> 
> I'm trying to push the limits and i saw that those have a crypto
> acceleration support:
> https://www.kernel.org/doc/Documentation/devicetree/bindings/crypto/marvell-cesa.txt
> 
> However with kernel 4.5.7 it does not really work and loading of
> marvel-cesa gives out error:
> [6.841862] marvell-cesa: probe of d0090000.crypto failed with error -22
> 
> :(
> 
> 
> I've only built latest Tor 0.2.9 with -O9 -mcpu=marvell-pj4
> -mtune=xscale and the Tor Relay is at
> https://atlas.torproject.org/#details/82C63C5D61D17557B7C5D0E7FDC545A2B6B6B7E0
> .
> 
> Does someone have experienced optimizing and tuning Tor specifically on
> ARM processors to understand how far those could be optimized?

Do you run two Tor instances per machine? On platforms like these (with two
or more cores, but each core being rather weak), you pretty much have to, if
you want good resource utilization. Pin the 1st instance to cores 0,1, the
second one to 2,3 using schedtool for better cache locality, and if you're
already using schedtool, might as well mark the Tor processes as SCHED_BATCH.

-- 
With respect,
Roman