[AIPS] [NRAO] AIPS Benchmarks: comments on SUN computers

Most recent change on $Date: 2002/08/14 18:03:29 $ (UT) [Athlon results]


Recent DDT Benchmark Results

SPARC ULTRA

Thanks to the courtesy, time and hard work of some Sun Engineers, NRAO has been able to perform the DDT benchmark on both an Ultra 1 (model 170) and Ultra 2 (dual processor, 182 MHz) SPARC. The Ultra 1 produced a best result of 6.3, and the Ultra 2 produced an AIPSMark(93) of 9.0. For the details, see AIPS Memo 91. On 25 February 1998 tests were made on three new Ultra models, the 5, 10, and 30. Sun engineers suggested that we assign an AIPS scratch-file only disk to the "disk" area called /tmp. This disk area is normally resident in memory rather than being cached to actual disk; it is equivalent to the swap area. This "trick" had a surprising effect in the results, substantially improving the real times and the cpu/real ratios (which were very low on the Ultras).

Unfortunately, this trick can be very dangerous since it can use up the swap space required to run programs. If that happens, the machine may halt. If a large swap space is configured in the system, this trick may be invoked in AIPS through assigning a disk area to /tmp, defining it to be scratch only, and then selecting it via the BADDISK adverb.

Here is an account of what happens in practice when this trick is used, from an actual user:

I'm using Solaris 5.6 with 15OCT98 AIPS and employing the "/tmp trick". Unfortunately, when any task creates a large temporary or scratch file on the /tmp disk, the machine becomes unresponsive to all users. Keystrokes are not echoed, window managers don't respond to mouse movements or clicks, etc. This can last up to a couple of minutes and with several AIPS users on the machine, makes things quite painful.

So users are advised strongly to be cautious before using the "/tmp trick";. The cost may not be worth it.

Chris Flatters has come up with some recommendations for Tuning Solaris for AIPS that appear to give most of the advantages of this "/tmp" trick without the risks mentioned above. People with memory-rich Sparc Ultra systems should take heed.

What about that SPARC Weitek Upgrade?

Henrietta May at The Australia Telescope National Facility has benchmarked a SPARC 1 with the 15JUL93 DDT before and after the upgrade. The results are quite interesting. The ATNF shares a site and facilities with the CSIRO Division of RadioPhysics.


Valid HTML 3.2!
[Powered by Apache!] Patrick P. Murphy