Author: Matt Taylor
Date: 19:40:30 02/24/03
Go up one level in this thread
On February 24, 2003 at 20:41:35, Robert Hyatt wrote: >On February 24, 2003 at 18:53:37, Matt Taylor wrote: > >>On February 24, 2003 at 18:06:14, Jeremiah Penery wrote: >> >>>I will respond to the stuff below, but the discussion is starting to get off >>>course. My assertion was that if Xeons suddenly added 50% to their clockrate >>>overnight that they would begin to eat into the 'server' markets. There are >>>plenty of applications in that space that are CPU bound, where that super fast >>>Xeon would fit nicely. Of course it would not take the entire market, or even >>>50% of the market. I never said it would. But I'd be willing to bet anything >>>that it would take _some_ of that market (5%, 10%, who knows?). That's all I >>>ever tried to claim in this particular thread. >>> >>> >>>On February 24, 2003 at 00:03:08, Robert Hyatt wrote: >>> >>>>On February 23, 2003 at 22:48:35, Jeremiah Penery wrote: >>>> >>>>>You seem to be ignoring that TPC-W has non-clustered x86 machines in the lead. >>>> >>>> >>>>Where? >>>> >>>>Didn't see a one that wasn't a NUMA-type box with each machine having its >>>>own I/O.... >>>> >>>>I may have overlooked something of course. >>> >>>I posted it a few messages up in this thread. But I overlooked something also, >>>in that every submitted result for TPC-W is an x86 machine. They're all listed >>>as non-cluster machines, up to 16 CPUs, but I don't know what their definition >>>of 'cluster' is. >>> >>> >>>I see something else interesting though. Top 10 TPC-C results for >>>non-clustered(*) machines look like this: >>> >>>1) 128 CPU Fujitsu SPARC64 GP 563MHz >>>2) 32 CPU Itanium2 1GHz >>>3) 32 CPU POWER4 1.3GHz >>>4) 64 CPU PA-RISC 8700 875MHz >>>5) <same as 3> >>>6) <same as 4> >>>7) <same as 2> >>>8) 32 CPU XeonMP 2GHz >>>9) 32 CPU Alpha 21264A 1001MHz >>>10) 48 CPU Sun SPARC64 GP 563MHz >>> >>>I'll isolate #s 8 and 9 here: >>> >>>8) >>>Total System Cost - 2,715,310 US $ >>>TPC-C Throughput - 234,325 >>>Price/Performance - 11.59 US $ >>>Availability Date - 03/31/03 >>>Database Manager - Microsoft SQL Server 2000 Enterprise Edition >>>Operating System - Microsoft Windows .NET Server 2003 Datacenter Edt. >>>Transaction Monitor - Microsoft COM+ >>> >>>9) >>>Total System Cost - 10,286,029 US $ >>>TPC-C Throughput - 230,533 >>>Price/Performance - 44.62 US $ >>>Availability Date - 07/30/01 >>>Database Manager - Oracle 9i Database Enterprise Edition >>>Operating System - Compaq Tru64 UNIX V5.1 >>>Transaction Monitor - Compaq DB Web Connector V1.1 >>> >>>How can such a number be explained? I would expect the Alpha machine to win by >>>a large margin, but it actually loses. >>> >>>(*) Again, I don't know how they define cluster. I am not aware of a Windows >>>version that has any kind of NUMA optimizations, however, which I think would be >>>necessary to get a very good score on this type of benchmark, if indeed the >>>machines are NUMA ones. >> >>"Operating System - Microsoft Windows .NET Server 2003 Datacenter Edt." >> >>Hello NUMA optimizations. >> >>I, too, am not sure why the Alpha is beaten by the 2 GHz XeonMP (though only by >>a narrow margin, <2%). I would say that comparing different databases is asking >>for trouble, but I have heard that Oracle is the fastest database software >>available. >> >>I believe all Intel systems with more than 16 CPUs are -definitely- NUMA. >>Supposedly they support 16P, but I have never seen one with that many >>processors. Most larger systems are built with a NUMA cluster of 4P nodes. >> >>-Matt > > >there are a couple of companies producing non-NUMA machines. IE Sequent in >Portland >Oregon once made a 32-way X86 box. Fully shared memory SMP platform... > >But they are not in the PC price range, starting at 1/2 mil and going up >_quickly_ as you >add processors. The largest I have seen is a 256-node NUMA system build with 4P Xeon nodes. That's 1,024 processors. I have no idea what it cost, but I'm sure it carries a hefty pricetag. -Matt
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.