[ale] sorta [OT] IO vs clock cycles

[Dow Hurst]

SGIs are engineered at the Origin 2000 level to have balanced bandwidth between the CPUs, memory, and internal bus.  They support about 650Mbps between all components.  This was part of the design for scaling to large symmetric parallel systems of up to 512 CPUs on one machine while maintaining performance.  You know the new itanium clusters from SGI will have similar performance scaling under Linux with the new code in their systems up to 64 CPUs.

For competition with RISC based processors, the Athlon and Pentium CISC based machines have to run a lot faster to get the same fpu.  One major reason that people still value the big hardware.  Just look up performance specs.  The RISC pipeline for FPU must be tremendously more efficient.  I don't know much about that stuff but the specs speak for themselves.  The price of commodity hardware offsets that performance by dropping the price on large clusters of Athlons.  But, in an equal cluster of expensive RISC chips versus less expensive CISC chips, the RISC will blow the CISC away for FPU.  This is why we are still using used SGIs over investing in a Athlon cluster.  Pricewise we can get the performance/price we need for the software we already own.  As we move into code we have the Source (Yes!) for, we will move toward a possible RISC based or Athlon based cluster depending on how much we can afford.  Biochemical computation biases us toward FPU, so my thoughts are skewed in that direction.

Please excuse my following rant:

For along time I have heard people say how Suns, HP, and all the other big hardware are great and why do we use SGI?  Also, I have heard people say "get a Linux cluster".  I am not against commodity Linux clusters since the price and OS are fun and cheap compared to the other choices.  I don't consider MicroSoft to even compete.  However, most sysadmins and IT people I talked with really have never seemed to understand the SGI Origin 2000/3000 benefits compared to the other big iron.  Hands down SGI runs rings around the the competition when you look at what you get if you have the bucks to afford the initial investment.  The scalability of every component, along with unreal flexibility with how those components can be portioned out make an Origin 2000/3000 level machine much more valuable in the long run than any other Unix big iron.  Only Cray really competed on all levels.  The numbers are staggering when you see that on a 1996 Origin 2000 with 256 CPUs you could 1024 SCSI controllers, 1Gb of RAM per CPU, each CPU and disk spindle could be portioned out to a single task, process, or user.  Realtime CPU, I/O, and disk spindle performance could be guaranteed for specific processes.  I don't know how many ethernet ports were allowed.  Your talking several million dollars in 1996 to set the thing up at that level.  But, the scalability of the architecture gave a business the chance to buy what it needed and just add to it as it grew it's needs.  No need for multiple servers or having to replace an entire box just to get a bigger box.  I know many will have different opinions, but I believe SGI showed the way on the infinitely scalable platform.  It is worth an examination as a good example of what Linux SMP could become once someone like HP or SGI ramps up the hardware toward the example of the Origin platform.  Now many companies have ramped up their scalable architecture to compete.  Any one have examples of high performance computers that could compete with top of the line SGI machines in terms of scalability and performance with true SMP?  Okay, enough ranting!!
Dow


>>> Mike Panetta <ahuitzot at mindspring.com> 11/09/02 01:46 AM >>>
Um, just to correct your post, PCI is 32 Bits and 64Bits, not 16 bits. 
It was never 16 bits.  ISA on the other hand is 16Bits.  And I think MCA
is too but I am not sure.

And with PCI its more complex then just how many bytes can you transfer
per bus clock.  First of all PCI multiplexes the address and data
together, so (and I may be incorrect here) the very first transfer on
the PCI bus is the address xfer.  After that you can do up to (I think)
512 data xfers (or whatever a cache line is on your CPU).  IE PCI
supports bursting, and you actually need to use it to get the full
bandwidth out of it.

Mike

On Fri, 2002-11-08 at 20:03, Larry Grenevitch wrote:
> To reply to the mainframe question.
> CPU speed is only one factor.  Several years ago I worked in a shop that sold 
> both AIX and SCO Unix boxes.  A 2 year old AIX box still beat out the latest 
> PC with on cache boards etc..  Why, IO, most PC boards are 16 bit with PCI 
> there is now 64 bit boards, on AIX and I am sure the mainframes you have 512 
> bit IO boards.  Or to put it another way, when your PC has gotten 16 bits [2 
> bytes of information] your mainframe transfered 512 bits or 64 bytes of 
> information.  And it is things like that all through out the system from IO 
> to memory.  That is why PCs are not really scallable, they just don't have 
> the base engineering.
> 



---
This message has been sent through the ALE general discussion list.
See http://www.ale.org/mailing-lists.shtml for more info. Problems should be 
sent to listmaster at ale dot org.



---
This message has been sent through the ALE general discussion list.
See http://www.ale.org/mailing-lists.shtml for more info. Problems should be
sent to listmaster at ale dot org.