[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: ssh-vs-rsh benchmark result

>>>>> "adam" == Adam C Powell, IV <Adam> writes:

    adam> But more importantly for us Beowulf users, who might run a job for
    adam> days and not care about a few seconds of startup time, I have been
    adam> under the impression that ssh encrypts all communications during the
    adam> session, which would use considerable CPU time in a parallel job with
    adam> lots of communication.  The more communication and faster the
    adam> connections (e.g. gigabit), the more CPU time would have to be devoted
    adam> to this task. Based on this understanding, I've been putting my
    adam> Beowulfs on private nets and using rsh.

This depends on how exactly the ssh/rsh connection is being used.  If it
is to be used as the pipeline for sending data, you are entirely
correct.  This is not necessarily the case; PVM can be set to use its
own communications setup, so that ssh is soley to initialize the
remote node message passing servers.  

As this thread again makes clear, one must be specific about the
particular type of local execution being run, and more details than
just "running on a beowulf" are of course necessary. 

(length of subprocesses/jobs, topology of the jobs of the executing
process, etc, etc)

    adam> Is my basic understanding sound, or does ssh send the session info, X
    adam> stuff, etc. in the clear?  Or does this not add significant
    adam> computation time, e.g. on the order of what the kernel uses to get the
    adam> packets out through the NIC?  Has anyone seen/done relevant benchmarks
    adam> which would test the magnitude of this effect?

It's a pretty hefty load for high i/o bandwidth communication (remote
X with java swing applications, for example :-).

You don't want to schlep large amounts of data through SSH; RSH and
PVM are better, and if you really care, it appears that RPC and direct
sockets are preferred (according to a protein structure predicting
friend of mine).


A.J. Rossini				Rsrch. Asst. Prof. of Biostatistics
U. of Washington Biostatistics		rossini@u.washington.edu	
FHCRC/SCHARP/HIV Vaccine Trials Net	rossini@scharp.org
-------------- http://software.biostat.washington.edu/ ----------------
FHCRC: M: 206-667-7025 (fax=4812)|Voicemail is pretty sketchy/use Email
UW:   Th: 206-543-1044 (fax=3286)|Change last 4 digits of phone to FAX
(my tuesday/wednesday/friday locations are completely unpredictable.)

Reply to: