SGI Performance Comparisons

Archiving a large number of small files to DAT (DDS1 vs. DDS3)

Last Change: 08/May/2008

DAT TEST 3: Archiving a large number of small files to DAT (DDS1 vs. DDS3).

This test measures DAT performance when archiving a large number of small files. It's difficult to construct a standard test for this, so my solution was to use the data in the /usr/share directory which is created after a default IRIX 6.5 installation. Even so, different systems end up with slight differences in the size of /usr/share - this is because of differences in default software selections during the installation process. Thus, for this test, it is the data transfer rate which is more important.

The command used to time this test, which was executed in a raw xterm from within the /usr/share directory, was:

   timex tar c .

In Table 24 below, the times are also included for reference, but do not use the times for comparisons between systems. O2 and Indigo2 GR3-Elan end up with 90MB in a default /usr/share, while Indy has its default /usr/share taking up 74MB. What matters is the aggregate data transfer rate.

                        DDS1 Python      DDS1 Python      DDS3 Sony
                       25601-XXX2.75    28388-XXX5.AC      SDT9000
                        rate (time)      rate (time)     rate (time)

O2 R5200SC 300MHz 1MB:                    238 (6:27)
O2 R5000SC/200:                                           819 (1:53)
Indigo2 R4400SC/250:     194 (07:56)                      819 (1:53)
Indy R4400SC/200:                                         455 (2:46)
Indy R4600PC/133:                                         326 (3:51)
Indy R4600PC/100:                                         356 (3:32)

      Table 24: DDS1 vs. DDS3 performance for
         archiving many small files to DAT.

The most interesting observations here are:

Indigo2 R4400SC/250 probably does well compared to Indy R4400SC/200 because of it's larger 2MB L2 and high clock speed, ie. the CPU can deal with the processing overhead more efficiently than lower-clocked CPUs which have a smaller L2. O2 R5000SC/200 performs faster than Indy R4400SC/200 probably because it has a better overall design.