Geoff O'Callaghan: openssh crypto cipher performance

It was mentioned to me that when transferring files on an internal network that by selecting a different cryptographic cipher you could improve the file transfer performance.  So, since I had a few spare minutes and elected not to scratch my bum I whipped up the following little script to test the theory.

I elected to scp a random ~700Mb file I affectionately called disc1.iso  (it was actually just random data, but you get the idea) to my localhost.  That is, I transferred the file from system A to system A.   I’m not interested in getting the highest possible speed with this test, i’m more interested in the relative performance of the ciphers.   Doing this creates a ‘relatively’ stable environment to conduct the comparisons.

I added my ssh key to allow myself to talk to myself - sort of like this blog really with the number of readers I have :-)  Then I did the following (a man ssh shows the valid ciphers for protocol 2)

for c in 3des-cbc aes128-cbc aes192-cbc aes256-cbc aes128-ctr aes192-ctr aes256-ctr \
         arcfour128 arcfour256 arcfour blowfish-cbc cast128-cbc ; \
         do for j in `seq 1 1`  ; \
          do /usr/bin/time -a -o results.txt -f "$c,$j,%E,%U,%S" scp -c $c disc1.iso localhost:tmp/ ;\
          done  ; \
         done &

This creates a results file which in my case looks like this :

3des-cbc,1,1:12.67,35.41,3.53
aes128-cbc,1,0:56.18,9.52,4.09
aes192-cbc,1,0:54.58,9.86,4.16
aes256-cbc,1,0:55.73,11.46,3.89
aes128-ctr,1,0:59.78,13.43,4.14
aes192-ctr,1,1:04.33,14.67,4.19
aes256-ctr,1,1:01.07,15.31,4.08
arcfour128,1,0:57.75,7.10,4.50
arcfour256,1,1:18.06,7.80,4.56
arcfour,1,0:59.32,7.05,4.60
blowfish-cbc,1,1:01.19,11.62,4.46
cast128-cbc,1,1:26.57,22.31,4.14

Now, according to the man page aes128-cbc is the default cipher for Protocol version 2 so if I use this as the baseline then the relative performance becomes  :

Based on those numbers I really wouldn’t bother trying to select a different cipher for the file transfer.

Note 1: This was performed on a run of the mill core 2 duo system running Ubuntu Hardy, you will possibly find that certain architectures have better results with certain ciphers possibly due to the instruction set being a better fit for a certain algorithm or in the case of higher end servers the availability and use of cryptographic hardware.

Note 2:  The seq 1 1 allows you to run the test multiple times, just change it to seq 1 10 to run each test 10 times.  I just did it once for the purposes of putting it in the blog.