Failed to hand off onionskin

Roger Dingledine arma at mit.edu
Tue Dec 16 23:44:58 UTC 2008


On Wed, Dec 17, 2008 at 12:28:37AM +0100, Mitar wrote:
> From time to time I am getting this warning:
> 
> Failed to hand off onionskin. Closing.
> Your computer is too slow to handle this many circuit creation
> requests! Please consider using the MaxAdvertisedBandwidth config
> option or choosing a more restricted exit policy.

This warning happens when you have more than 100 create cells queued
for processing, so presumably your CPU isn't able to keep up with them.

> I have been monitoring the system and while it is true that sometimes
> it tops one processor, it occupies most of the time just 50 % of one
> processor. I have also configured Tor daemon to use two threads so
> even if it tops one it could still switch to another. But it rarely
> passes 100 % (that is, it rarely really uses two processors). The
> system as a whole has also not topped its CPU power. And while load
> does not seem to be so high I get at the same time this errors. Is
> there some other system bound which would be causing this and not CPU?
> Are there some other performance tweaks I could try?

Well, step one is to set
NumCPUs 2
but I think you already did that.

Assuming you're referring to the relay "Arlequin", this is a very fast
relay, so you will be running up against various constraints that not
many people hit.

We have a todo item to make Tor better on using multiple CPUs. See
"Handle multi-core cpus better" at
https://svn.torproject.org/svn/tor/trunk/doc/TODO.021

But the case of decrypting onion requests (aka create cells) is actually
the part where we do best already at using multiple CPUs.

> I am using Tor 0.2.0.31 (r16744) on FreeBSD amd64. Should I maybe
> upgrade to devel version?

An amd64 cpu is a good start, since Tor's crypto is much faster on 64-bit.

How much ram is Tor using? If it's a lot, you might consider the -alpha
versions, as on some platforms they use much less ram.

--Roger



More information about the tor-talk mailing list