Re: Wildfire/Oracle/NUMA/QBB affinity on OpenVms/Oracle73

From: Rob Young <young_r_at_eisner.encompasserve.org>
Date: 6 Feb 2001 11:50:10 -0500
Message-ID: <Zk4WJRQkfcLF@eisner.encompasserve.org>

In article <3a7fb92b.2226779_at_news-server>, nsouto_at_nsw.bigpond.net.au.nospam (Nuno Souto) writes:

> With new releases for that specific h/w, it's possible the problem
> will go away, although I doubt it will be a complete solution.
>

	Don't overlook hardware helping out too.  Some people get
	antsy when futures are trotted out.  But some folks make
	decisions based on futures (Sandia/Celera, Los Alamos, European
	SuperComputer Centre) so we shouldn't be that nervous.

	This afternoon, the following presentation takes place:

http://www.isscc.org/isscc/2001/ap/ap/AP_forWeb_Nov16.pdf

15.6 A 1.2 GHz Alpha Microprocessor with 44.8 GB/s Chip

Pin Bandwidth

Jain, et al. Feb 6. 4:15 p.m. Compaq Computer Corporation, Shrewsbury, MA

A 4th generation Alpha microprocessor running at 1.2 GHz delivers up to 44.8 GB/s chip pin bandwidth and dissipates 125W at 1.5V. It contains a 1.75MB 2nd level write-back-cache, two memory controllers supporting 8 Rambus(tm) channels running at 800 MB/s, four 6.4 GB/s inter-processor communications ports, and a seperate IO port capable of 6.4 GB/s. The chip measures 21.1x18.8 mm2 and contains 130M transistors.

	What interests me more than bandwidth is latency.  Latency is
	the issue (as everyone has bandwidth or soon to, i.e. Power4,
	Ultra III) now.  From what we see, latency gets much better
	with EV7, this link is no longer there:

http://www.alphapowered.com/alpha21364.ppt

	But if it was there you would notice that local latency is 
	"30 ns CAS latency pin to pin" (slide 17) and L2 latency is 
	"12 ns load to use" (slide 16) with "15 ns processor to processor
	latency" (slide 18, i.e. remote memory routing) so it *appears* if the
	memory is two hops away , you may be looking at < 150 ns memory
	access if the page is open (sure, add a few dozen nanoseconds for 
	routing , whatever).

	Point is latency for Alpha gets MUCH better and NUMA *should*
	become less of an issue for future Alpha hardware.  Perhaps they
	talk more about latency this afternoon.
	
				Rob

Received on Tue Feb 06 2001 - 10:50:10 CST