AMBER Archive (2003)

Subject: Re: AMBER: PMEMD Performance on Beowulf systems

From: Robert Duke (rduke_at_email.unc.edu)
Date: Mon Dec 22 2003 - 10:44:47 CST

David -
Thanks much for the specific inputs on good vs. bad cluster hardware. I
have observed differences between cheap and more expensive hardware, but not
systematically. In general, I would trust server grade setups, including
switches, more than the average desktop pc, which can be a piece of junk
from an i/o perspective. However, specific recommendations are more useful
than general guidelines. PMEMD is designed with computation and i/o overlap
in mind; but not all hardware will take good advantage of it.
Regards - Bob

----- Original Message -----
From: "David E. Konerding" <dekonerding_at_lbl.gov>
To: <amber_at_scripps.edu>
Sent: Monday, December 22, 2003 10:53 AM
Subject: Re: AMBER: PMEMD Performance on Beowulf systems

> Viktor Hornak wrote:
>
> >
> > A7M266-D Dual Socket A Motherboard (AMD 762 Chipset). It has 3 32bit
> > 33MHz PCI slots and 2 64/32bit 66/33MHz PCI slots. To get a noticable
> > speedup in networking, the gigabit card (Intel Pro1000) needs to be
> > placed into 64bit 66MHz PCI slot.
> >
>
> A couple more notes:
>
> 1) Make sure the slot is configured in BIOS for the maximum speed. For
> example, our latest board has a 133MHz PCI slot, but the BIOS sets the
> default to 100MHz
> 2) Specific chips in the Intel gigabit series are better than others.
> For example, from our local networking guru:
>
> >In general, copper GigE (specially Intel onboard 84540EM) is cheap
chipset.
> >That is, they do not work well -- chews to much CPU. Try to avoid Intel
> >82540/82541 GigE NICs.
> >Intel 82545EM copper GigE is the only one that we found works well.
> >
> >
> >Fiber NICs are usually better, but not always. Becareful what you want.
> >
>
> From what I can tell in the code (I'm not an MPI expert, just enough to
> get into trouble) PMEMD makes more use of overlapping
> communication and computation than AMBER (Robert & AMBER gurus let me
> know if this is wrong). This is interesting
> because, for example, the 82545EM can sustain gigabit speeds in our
> machine (dual Xeon 2.8GHz on SuperMicro
>
> X5DPL-iGM motherboard) while leaving 75% of one CPU and 100% of the other
free.
>
> The other interesting thing I noticed is that HyperThreading gave an
unexpected benefit- if I run PMEMD or AMBER
> with np=4 on a 2 processor (+2 HT processors) machine, I actually get a
mild speedup relative to just 2 cpus. This means, for example, that single
> SMP boxes can run ~1ns sims in reasonable time, so we're not really using
the gigabit interconnect much at all.
>
> Also, at Supercomputing some data about 10GbE was released. There are
some suggestions 10GbE made a huge difference in latency relative to 1GbE,
> meaning there may be plain 10GbE-based interconnects that are competitive
with the proprietary interconnects, but hopefully at a lower
> cost. That, combined with Infiniband interfaces on the north bridge
should definitely help a lot.
>
> Dave
>
>
>
>
> -----------------------------------------------------------------------
> The AMBER Mail Reflector
> To post, send mail to amber_at_scripps.edu
> To unsubscribe, send "unsubscribe amber" to majordomo_at_scripps.edu
>
>

-----------------------------------------------------------------------
The AMBER Mail Reflector
To post, send mail to amber_at_scripps.edu
To unsubscribe, send "unsubscribe amber" to majordomo_at_scripps.edu