Was [Gllug] Best option for a lot of compute power

[Andy Farnsworth]

-----Original Message-----
From: Luke Hopkins
Sent: Tuesday, June 08, 2004 6:35 PM
To: 'Daniel P. Berrange'; 'Greater London Linux Users Group'
Subject: RE: Was [Gllug] Best option for a lot of compute power


Following on from this (a little more OT), whats the consensus on
performance differences between Ultra 320 SCSI & SATA 150? I realise
these are bus speeds, but how about running 2 disks per bus, could these
speeds be reached?

-----End Original Message-----

Luke,

  While SCSI allow multiple disks per bus, Serial ATA is a point to
point protocol (i.e. one cable, one disk) and while this can be changed
by adding a controller (i.e. computer SATA -> External Device with SATA
controllers -> Multiple drives) this is not usually the case.  This
means that all the drives on the SCSI bus share the 320 Mbit bandwidth
while with SATA each drive has access to the full bandwidth (provided
the controller can handle it).  This should indicate that a SATA RAID
has more overall bandwidth than Ultra SCSI 320 if you have 3 or more
drives.  That said, you also have to consider the latency of the drives
in question as well as the bandwidth and SATA drives are limited to
rotation speeds of 7200 (though there may now be 10,000) RPM while SCSI
have reached 15,000 RPM giving a much better latency.  This may not seem
like much but if you do the math and then compare it to the speed of
your processor you will quickly realize that latency is a huge hit when
it comes to performance.

10,000 = 1/10 ms (milisecond or 1/1000 of a second) for one rotation

Average case is 1/2 rotation to get to your data gives 1/20 ms average
latency.

15,000 = 1/15 ms for one rotation

Average case is 1/30 ms average latency.

Remember that your computer processor runs something like 2 Ghz which is
2 BILLION cycles per second.  Thus in 1/20,000 second the processor has
sat idle for 2,000,000,000/20,000 or 100,000 cycles and in 1/30,000
second the processor sits idle for only 2,000,000,000/30,000 or 66,667
cycles.  This is a huge savings in time and get's more pronounced the
faster CPUs get.

See http://flash.stonedoor.com/Latency_Impact.html for a full chart.

Andy


-- 
Gllug mailing list  -  Gllug at gllug.org.uk
http://lists.gllug.org.uk/mailman/listinfo/gllug