On 10/25/10 4:39 PM, Dan Pritts wrote: > > On Oct 25, 2010, at 4:35 PM, Steve Simmons wrote: >> A couple of years ago we did some performance testing on different ssh >> encryption algorithms for bulk transfer to a single host. Forcing use of >> blowfish gave us the fastest throughput. This was at least partially a CPU >> issue; there were upwards of 10 machines dumping data via ssh to a single >> host with the equiv of 'tar cf - ... | ssh -c blowfish hostname tar xpf - -C >> /dirname'. That's close enough to Dan's suggestion that I'd expect blowfish >> to be a win for you. > >> Assuming you don't mind the data being viewed in-transit, get the Pittsburgh >> Supercomputer ssh version. It has switches that allow you to encrypt the >> authentication steps but sends the data in plaintext. If you do lots of >> simultaneous instances, CPU is drastically reduced or removed as bottleneck. > > It also has an encryption algorithm that utilizes multiple cores > > More importantly, it also fixes the totally broken nature of ssh's internal > windowing.
Thanks for the workaround ideas... With blowfish, I'm getting over 150Mb/sec instead of 30Mb/sec. I'm still stuck on why the standard moves are so slow... UDP is probably out of the equation, as I was able to do an iperf test at 900+MB/sec between two AFS servers once I had the bandwidth option set correctly (rather that defaulting). Even using machines on the same campus, the same switch, I never get them to break about 30Mb/sec, not even with multiple vos moves going in parallel. It could very well be firewall-related, it could be causing fragmenting and causing rx to retransmit. What kind of throughput SHOULD I be seeing with Gigabit (but no Jumbo frames, we had even more severe slowness due to fragmentation to another network when we had that turned on... it will be revisited) connections? Has anyone else run into this severe a slowdown in moving large volumes from server to server? Thanks again, Chris -- Eric Chris Garrison | Principal Mass Storage Specialist [email protected] | Indiana University - Research Storage W: 317-278-1207 M: 317-250-8649 | Jabber IM: [email protected] _______________________________________________ OpenAFS-info mailing list [email protected] https://lists.openafs.org/mailman/listinfo/openafs-info
