[strongSwan] Latency and throughput
sangdrax8 at gmail.com
Mon Dec 14 19:36:05 CET 2015
On Fri, Dec 11, 2015 at 4:17 PM sangdrax8 <sangdrax8 at gmail.com> wrote:
> I am having some problems with throughput in situations with high BPD. If
> anyone has some suggestions it would be much appreciated.
> My setup is currently as follows:
> 2 Ubuntu boxes running Strongswan 5.3.2 and kernel 3.13.0-71-generic. I
> am using IKEv2 (don't think that matters), ESP in tunnel mode, with
> aes256-modp3072-esn, and I set the replay_window=0. I connect both devices
> through a third box that I use to adjust the latency between the devices.
> If I do not increase latency, I can iperf3 between both servers at 900's
> Mbps, both with and with out encryption. The boxes don't show any
> significant load, so I wouldn't expect to have hardware issues after the
> latency between them increases. As I increase latency, my throughput
> starts to go down if ipsec is used. With latency of just 30ms, I see a 1
> minute test push near 600 Mbps. If I simply turn off ipsec, and run the
> same test, I hit 900's with in the first 2-3 seconds and stay there.
> This gets much worse as I jump up to 100ms. With some TCP tuning, my
> unencrypted transfer reaches the 900's with in about 10 seconds. My
> encrypted transfer just hovers around 100Mbps. I only achieved this after
> reading a previous post about the replay_window. With replay window on I
> wasn't able to push even 10Mbps.
> This feels like the replay_window, or the TCP tuning with tcp_rmem/wmem
> that I needed to do to account for the extra packets that are in flight
> over the link with such high latency and bandwidth. ESP isn't TCP, so I
> have increased wmem_default, wmem_max, rmem_default, and rmem_max to all be
> the max value I used when doing my TCP tuning, but I still get the 100Mbps
> while encrypting.
> Any suggestions/pointers as to other tuning steps that might be needed to
> ensure packets are sent at the higher rate I know this setup can achieve?
> Is there still some replay_window type delays taking place here that I have
> Thank you.
Just in case someone else sees this, it seems my issue was with the tc
rules providing the latency between my two machines, and not related to the
encryption. Not sure why it didn't show up when encryption was off, but
increasing the limits lets me run my tests at higher latencies with
encryption turned on.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Users