3/95 Direct Attach Disk Adhoc Minutes

Radek Aster raster at cherokee.asd.sgi.com
Thu Mar 9 15:55:00 PST 1995

>>>>> "Kurt" == Kurt Chan <kc at core.rose.hp.com> writes:

Kurt>      [Note: Radek's simulations show that for the I/O per second
Kurt> benchmark, configurations with Login_BB_Credit = 0 at both
Kurt> Initiator and Target are virtually INDISTINGUISHABLE in
Kurt> performance from any other combination of Login_BB_Credit > 0
Kurt> for loops with less than 70-80 drives (where the best
Kurt> combination peaks at 100 drives).  In the throughput benchmark,
Kurt> Login_BB_Credit = 0 is virtually IDENTICAL to the performance of
Kurt> Login_BB_Credit > 0 if Available_BB_Credit is 2 or 3.]

One has to be careful when making general interpretations based around
the number of devices. The simulations were done assuming
Seagate-Barracuda class drives; i.e. 5400 RPM, 15 mSec Max Seek,
etc. which implies 73 random 2K IOPS per spindle.

As drives become faster, (7200 RPM, 120 max random IOPS per spindle)
the number of devices one can stick on a loop before saturation sets
in will come down. When this happens, in the future, non-zero
login_BB_Credit WILL become a more important factor in determining
connectivity vs. performance tradeoffs.

 Radek Aster                            @       | "If humans are so evolved,
 mailto:raster at sgi.com      ___~*:    _ \    _  |  how come they don't have
 mailto:raster at netcom.com  '/|~~\    (*)-^+-(*) |  opposable thumbs on their
                                                |  feet ?"
 http://reality.sgi.com/employees/raster/       | 

More information about the T10 mailing list