EMME/2 under SCO UNIX

As mentioned in the editorial, EMME/2 has now been ported to the SCO Unix operating system on INTEL machines. It is actually being tested internally at INRO, and should be released officially in a few months.

INTEL 486 based DOS machines rank very well on CPU intensive tasks, such as the EMME/2 fixed demand auto and transit assignments. On the other hand, disk intensive tasks, such as the variable demand auto assignment and matrix calculations, are very slow. By comparing two identical computers, one running under DOS and the other under Unix, we can now verify if it is the hardware implementation or the operating system that is the limiting factor. Is it worthwhile to transform a DOS machine into a Unix machine to gain speed? We will try here to give a few test results to help the decision making.

We compared two identical computers: two Compaq 486/25 with 6MB of main memory, both equipped with a Trident super VGA graphic card. The two FORTRAN compilers used to create the EMME/2 binaries are from the same company. Four benchmarks were used for the comparison:

• the fixed demand auto assignment: this is our usual CPU benchmark;
• DBENCH: a disk oriented benchmark which consists of a simple calculation with standard Winnipeg matrices: mf6 = mf1 - mf2' + mf3 - mf4';
• the variable demand auto assignment: this type of assignment is both CPU and disk intensive;
• GBENCH: our usual graphic benchmark (see EMME/2 News of June 1991).

The following table gives the final results, along with results obtained for other Unix workstations.

Fixed Demand Variable Demand

Computer Auto Assignment DBENCH Auto Assignment GBENCH

1 it. Total 1 it. Total

I486 DOS 14.6 174.8 127.4 69.2 852.2 8.8

I486 DOS with cache 13.3 149.7 40.0 27.7 420.3 8.7

I486 SCO 11.3 122.3 16.2 16.3 175.9 10.0

SUN SPARC SLC 15.2 167.0 14.9 19.0 209.3 9.1

SUN SPARC IPX 6.7 74.1 10.1 8.8 90.7 3.6

IBM 6000/320h 5.4 59.0 8.8 6.9 75.6 3.3

HP 9000/720 3.0 32.7 5.7 3.8 41.9 1.8

SUN SPARC SERV 330 11.1 121.8 15.8 14.6 160.8 6.2

Clearly, DOS needs help when it comes to disk access. Installing a disk cache gives a 50% improvement, and once SCO is installed, the Compaq catches up with other Unix workstations having a similar CPU performance. With 33 and 50 Mhz I486, results ought to be even better.

Three points are worth mentioning concerning these results:

• The first results obtained on the DOS machine with the DBENCH and variable demand macros were 30 to 40 percent slower than those given here. After examining the hard disk with the Norton Utility program, we discovered that the file containing the EMME/2 data bank was highly fragmented. It was scattered into more than 20 pieces all over the disk. Thus, by using a utility program to reallocate the disk blocks and have the data bank stored on contiguous blocks, we were able to markedly reduce the benchmark times. On DOS machines it is worthwhile to check for disk fragmentation once a data bank is created, especially if the disk is near capacity. Disk caching and defragmenting is automatic on most Unix systems, but it is the user responsibility on DOS systems.
• To run EMME/2 under SCO Unix, one needs to have the SCO Unix operating system and the SCO Xsight X Window system. Both can be purchased separately or bundled with the Open Desktop package.
• With 6MB of main memory on the machine running SCO, everything went very well in text mode. Once the X Window system was started, the time to load a program (e.g. go from one module to another) and the mouse response time began to be noticeably long. The X Window system needs a lot of memory and 8MB would be a bare minimum to install. 12MB would be preferable.

If your application is power hungry and you have an INTEL machine lying around, installing SCO might be an inexpensive solution to consider. If you can't stand living without DOS, note that SCO allows one to run DOS and Unix programs in parallel or use the station as a pure DOS machine.