Richard Maher wrote:
> 1) I saw 10 Gigabit NIC sup****t scheduled for VMS 8.3 (Depending on
which
> slide you look at it says "Integrity Servers Only"). Now 10x what a lot
of
> people are using for a cluster-interconnect at the moment sounds pretty
> s__t-hot to me! Especially if you're moving big lock-trees around the
> cluster. Given that this functionality is less than a year away, surely
some
> performance figures or at least anecdotal evidence should be available?
One way to estimate the impact is to see how Gigabit Ethernet compared
with Fast Ethernet. Bandwidth went up by close to 10x; latency went from
about 240 microseconds for a round-trip lock conversion request with
Fast Ethernet to about 200 microseconds on Gigabit Ethernet, as measured
on a typical Alpha box of the recent past (Wildfire or ES40). So don't
expect a 10x latency improvement.
> I mean, if I was an Rdb engineer that had used p___-poor DLM performance
as
> the rationale for sticking all of my R&D eggs in the stand-alone
single-node
> basket, then I'd be interested in what's happening with this. Right?
"But
> it's not the bandwidth, it's the latency that gets ya." Well that brings
me
> to the next slide. . .
Rdb sup****ts Row Cache in Galaxy Shared Memory between multiple nodes.
> 2) Next Generation Low-Latency Interconnects Post 8.3 (Integrity Servers
> Only)
>
> Am I the only person getting their jollies out of this or what?
Nope. I also find this exciting.
Potential candidate technologies one would naturally want to look at
could include Infiniband and RDMA/iWARP.
Infiniband promises low latency, but hasn't really taken off much in the
industry yet, and hardware is expensive. Some initial proponents have
subsequently backed out (like Intel). A lot of people are waiting to see
how this turns out.
RDMA/iWARP looks to have broad potential industry sup****t, and will
quite possibly be built into commodity Ethernet adapters. Latency
wouldn't be quite as low, but price would be low and price/performance
very good.
Should be interesting. In any case, I'm know VMS Engineering has its
finger on the pulse of the technologies available in the marketplace,
and will provide a quality solution with the best interests of the
customers in mind.
> Will there be a special limit on the distances between nodes for this
stuff
> to work? (Like memory channel) Can you have a Disaster Tolerant Low
Latency
> Cluster?
Infiniband has some fairly low distance limitations, unless you include
an Infiniband router, and that's for IP traffic.
For anything Ethernet-based I don't expect distance limitations. Other
than how far you can drive light over fiber, there's no inherent
distance limit in Gigabit Ethernet today, for example. But of course the
longer your inter-site distance, the more delay there is due to the
speed of light over the distance, which mitigates against low latency.
---
Looks likely I'll be doing a hands-on workshop at HP Technology Forum on
Long-Distance OpenVMS Clusters. We'll explore some of the impacts of
long distances on performance in that workshop, for folks who are
interested.
|
