[Tor-BSD] attic: randomness exhaustion
Richard Johnson
rdump at river.com
Mon Mar 16 13:22:15 EDT 2015
On 2015-03-15 02:26, Libertas wrote:
> That said, it seems that updating to Tor 0.2.6.3 has significantly
> increased my throughput.
I cleaned up our state limit blockage by boosting them a while ago and noticed
a slight throughput increase. Now I've noted an additional 1.7x throughput
increase about a week after upgrading to 0.2.6.3-alpha. (I don't know that
0.2.6.3-alpha is the sole cause of the increase. I will switch up to
0.2.6.4-rc this week.) Numbers from nearly 3 weeks on 0.2.6.3-alpha are below
for comparison.
There are 6 total relays (1 exit, 2 early in their lifecycle, 3 older
middle/guards) on our 2-CPU (AMD Opteron 2435, 2.6GHz, non-AESNI) 12-core
OpenBSD 5.6-stable system running tor 0.2.6.3-alpha. The currently highest
bandwidth relay is averaging 1.7GB/sec per arm, perhaps because it peaked over
6GB/sec for a few days [0]. Average throughput achieved by all 6 together is
now 4.6MB/sec per arm, up from 2.6MB/sec on 0.2.5.10.
As that throughput grows, we've started peaking over 14k current state entries
in pf for Tor traffic [1]. The state entry count would be even higher but for
"set optimization aggressive" in pf.conf expiring TCP leavings faster than the
norm (which naturally raises the state-mismatch events as later client packets
arrive).
One isolated congestion event showed with a filled ifq two weeks ago, even
with ...ifq.maxlen boosted to 1024, but otherwise no ifq problems [2]. As
well, our pf state table allocation is large enough that we're not running out
of memory for states [1].
Richard
-------
[0]
https://globe.torproject.org/#/relay/1CC39E06101B0DBFA103A18A2032C4A0FE0503C8
https://atlas.torproject.org/#details/1CC39E06101B0DBFA103A18A2032C4A0FE0503C8
[1] pfctl -si output from a near peak time:
State Table Total Rate
current entries 14226
searches 10297827608 9138.3/s
inserts 34574653 48.7/s
removals 34560427 48.7/s
Counters
match 51758879 59.9/s
bad-offset 0 0.0/s
fragment 968 0.0/s
short 845 0.0/s
normalize 200 0.0/s
memory 0 0.0/s
bad-timestamp 0 0.0/s
congestion 2023 0.0/s
ip-option 0 0.0/s
proto-cksum 118 0.0/s
state-mismatch 708205 0.6/s
state-insert 0 0.0/s
state-limit 0 0.0/s
src-limit 0 0.0/s
synproxy 0 0.0/s
translate 0 0.0/s
[2] select sysctl output:
kern.nfiles=13058
net.inet.ip.ifq.maxlen=1024
net.inet.ip.ifq.drops=3220
kern.netlivelocks=76
More information about the Tor-BSD
mailing list