Contributed by Janne Johansson on from the new mips on the clock dept.
In this commit, visa@ submitted code (disabled for now) to use built-in acceleration on octeon CPUs, much like AESNI for x86s.
I decided to test tcpbench(1) and IPsec, before and after updating and enabling the octcrypto(4) driver.
I didn't capture detailed perf stats from before the update, I had heard someone say that Edgerouter Lite boxes would only do some 6MBit/s over ipsec, so I set up a really simple ipsec.conf with ike esp from A to B leading to a policy of
esp tunnel from A to B spi 0xdeadbeef auth hmac-sha2-256 enc aesgoing from one ERL to another (I collect octeons, so I have a bunch to test with) and let tcpbench run for a while on it. My numbers hovered around 7Mbit/s, which coincided with what I've heard, and also that most of the CPU gets used while doing it.
Then I edited /sys/arch/octeon/conf/GENERIC, removed the # from octcrypto0 at mainbus0 and recompiled. Booted into the new kernel and got a octcrypto0 line in dmesg, and it was time to rock the ipsec tunnel again. The crypto algorithm and HMAC used by default on ipsec coincides nicely with the list of accelerated functions provided by the driver.
Before we get to tunnel traffic numbers, just one quick look at what systat pigs says while the ipsec is running at full steam:
PID USER NAME CPU 20\ 40\ 60\ 80\ 100\ 58917 root crypto 52.25 ################# 42636 root softnet 42.48 ############## (idle) 29.74 ######### 1059 root tcpbench 24.22 ####### 67777 root crynlk 19.58 ######So this indicates that the load from doing ipsec and generating the traffic is somewhat nicely evened out over the two cores in the Edgerouter, and there's even some CPU left unused, which means I can actually ssh into it and have it usable. I have had it running for almost 2 days now, moving some 2.1TB over the tunnel.
Now for the new and improved performance numbers:
204452123 4740752 37.402 100.00% Conn: 1 Mbps: 37.402 Peak Mbps: 58.870 Avg Mbps: 37.402 204453149 4692968 36.628 100.00% Conn: 1 Mbps: 36.628 Peak Mbps: 58.870 Avg Mbps: 36.628 204454167 5405552 42.480 100.00% Conn: 1 Mbps: 42.480 Peak Mbps: 58.870 Avg Mbps: 42.480 204455188 5202496 40.804 100.00% Conn: 1 Mbps: 40.804 Peak Mbps: 58.870 Avg Mbps: 40.804 204456194 5062208 40.256 100.00% Conn: 1 Mbps: 40.256 Peak Mbps: 58.870 Avg Mbps: 40.256
The tcpbench numbers fluctuate up and down a bit, but the output is nice enough to actually keep tabs on the peak values. Peaking to 58.8MBit/s! Of course, as you can see, the average is lower but nice anyhow.
A manyfold increase in performance, which is good enough in itself, but also moves the throughput from a speed that would make a poor but cheap gateway to something actually useful and decent for many home network speeds. Biggest problem after this gets enabled will be that my options to buy cheap used ERLs diminish.
(Comments are closed)
By Peter J. Philipp (pjp) nospam@solarscale.de on http://centroid.eu
This is incredible! Thanks visa! The question I have is, why is this disabled? Is there more work to be done? Can't wait to put this on both my octeons...I'm aiming for some weeks down the line.
-peter
Comments
By Janne Johansson (jj) jj@stacken.kth.se on http://www.inet6.se
Must be because it is super unstable. Send your octeons to me instead. 8^D