Terrible hme throughput

[Gary Palmer]

On Wed, Sep 27, 2006 at 02:19:16PM -0700, Frank Jahnke wrote:
> fj>> using scp on a large file are about 1.3MB/s.
> >
> > try an scp to localhost.  
> >
> > 1MB/s is about what I get with 500MHz UltraSPARC IIe.  The scp built
> > into Solaris (compiled with Studio or Forte or something I assume) is
> > a little over twice as fast on a 440MHz UltraSPARC IIi.  sorry I can't
> > compare exactly the same hardware, but I'm assuming the compiler is
> > probably the difference.
> 
> Well, I get about 0.9MB/s using scp to copy from the original file to 
> localhost.   The overhead on both ends of the transfer undoubtedly explains 
> the (slightly) lower transfer rates. 
> 
> *Sheesh*
> 
> OTOH, I recall that people are using this level of hardware as routers, and 
> I recall them getting about 30Mb/s (or 4MB/s) through them.  Is there really 
> that much overhead with scp?

>From memory, back in the dim and murky past when I was a Solaris admin,
I heard co-workers comment that the math involved in SSL/SSH encryption
was not particularly nice to SPARC/UltraSPARC CPUs.  Or rather, the
CPUs were not particularly fast when faced with that math.  End result -
booming business for SSL offloaders.

I've seen an Ultra II push > 20MBit/sec with Firewall 1 loaded doing
hme<->hme routing, so I strongly suspect its related to ssh/scp/sftp