Hot Diggety! Richard Sims was rumored to have written:
>
> I would start with the netstat inspection outlined in the "Backup
> performance" topic of ADSM QuickFacts, which will quickly assess how
> network flow is doing in the actual event.  TCP window size may be
> involved in what you are seeing, known to be a particular factor over
> long distance communication.

I regularly push up to 980 Mbps sustained over a nearly 3,600 mile
(~5800 km) WAN, but usually about 175-200 Mbps for a typical workload.

Richard is correct: TCP window size tuning is *vital* for a high
bandwidth WAN. Don't leave home without it correctly tuned. :-)

For instance, according to Sun support working one of our cases, Solaris
is tuned for LAN performance out of the box.

It's not a lot more to tune it for a WAN, but it does require some
careful notes, theory, measurement, and trial-and-error.

We also got a major boost in performance merely by enabling jumbo frames
on each end. A smaller boost from using certain tunables to push the
NICs to their absolute maximum on Solaris.

I haven't done AIX perf tuning in so long that I'm a little rusty now, alas.

But for AIX, one tunes sliding windows by adjusting these 'no' options:

        - tcp_sendspace
        - tcp_recvspace

There's a few more 'no' options... rfc1323 is another key one to enable.

http://www.performancewiki.com/aix-tuning.html

Regardless of the sliding windows values chosen, strongly suggest
testing and retesting with various (power-of-2) values until the optimal
value is seen. E.g. 1024, 2048, ..., 65536, etc.

Also, the original poster may want to check with the network folks
(and/or upstream provider if appropriate) to make sure there's no
particular segment along the end-to-end network path that is of small
bandwidth. In my case, it's all 'in-house' so that makes it easier.

E.g. you're not going to see gigabit performance end-to-end if you have,
say, a DS3 (45 Mbps) circuit somewhere in between.

Also, before making changes, record before/after values as well as
results of measurements. Nothing worse than getting muddled in the perf
tuning process due to lack of careful record-keeping.

Some food for thought.

-Dan