Here are the results of the tests:

A) 1 to 4 requestors, 1CPU, no compression

The K-10000 10BaseT and S-70000 10BaseT show similar performance, slightly better for the S-70000. The S-70000 shows much lower CPU usage than the K-10000 (see CPU consumption figures, later);

• For 10BaseT, adding requestors allow the doubling of the throughput rate (when comparing the 1-processor to the 4-processor);
• The 100BaseT scales very well, going from 600KB/Sec for a processor requestor to over 1530 KB/Sec with 4 processorsrequestors;
• ATM scores very well for 1 requestor processor (over 1600 KB/Sec), stays the same for 2 processors, drops at three requestor processors and goes up a little for 4 requestor processors. The drop in performance can not be explained at this point. The same behavior has been observed in the 2 to 8 series. We think it is a subtle interaction between the software (ADSM Requestor, TCP/IP driver) and the ATM hardware. This will require future investigation.
• The following table gives the approximate CPU consumption for the various transfers:

# Requestors	K-1K-Eth	S-7K-Eth	S-7K-ATM	S-7K-FastE
1	43%	16%	50%	22%
2	70%	25%	47%	40%
3	70%	29%	50%	47%
4	77%	33%	45%	62%

B) 1 to 4 requestors, 1CPU, compression     Data compression increases greatly the transfer rate if sufficient CPU is available. The figures for K-10000 are missing from the test but the curve should be similar to the one for the S-70000 but clipping will probably occur starting at 3 requestors, because of CPU saturation which will occur much earlier on the K-10000; Of course, the compression ratio will vary depending on the data being compressed. We were achieving a compression ratio of 2.5:1 to 3:1; Over 10BaseT and S-70000, transfer rates of up to 1,500KB/Sec (1,5MB/Sec) have been observed. The best achiever is 100BaseT, with transfer rates in excess of 2,200KB/Sec (2,2MB/Sec). The following table gives the approximate CPU consumption for the various transfers:   # ProcessorsRequestors K-1K-Eth S-7K-Eth S-7K-ATM S-7K-FastE 1 35% 42% 50% 2 75% 75% 80% 3 85% 95% 4 90% 99%

C. 2 to 8 requestors, 2 CPUs, no compression       In all cases, adding more requestorsprocessors to the job will increase the throughput rate. It is especially true on fast machines (S-70000) using fast medium (100BaseT or ATM); We observe the same strange behavior of ATM with decreasing performance from 2 to 4 to 6 requestorsprocessors. But strangely, with 8 requestorsprocessors, the transfer rate goes up very sharply! The figures obtained for 10BaseT show that the medium is being saturated (approaching the theoretical capacity of the medium) and adding requestorsprocessors will only increase marginally the transfer rate; The following table gives the approximate CPU consumption for the various transfers: # ProcessReqs. CPU 0 CPU 1 CPU 0 CPU 1 CPU 0 CPU 1 CPU 0 CPU 1 K-1K-Eth K-1K-Eth S-7K-Eth S-7K-Eth S-7K-ATM S-7K-ATM S-7K-FastE S-7K-FastE 2 50% 12% 21% 6% 27% 35% 29% 12% 4 60% 15% 26% 8% 27% 35% 38% 17% 6 55% 15% 27% 9% 25% 34% 8 57% 15% 28% 9% 27% 39% 46% 19% For most of the medium, the behavior is the same: the CPU usage of CPU 0 (the one that contains the stack) is much higher than CPU 1 that runs only the BCOM requestors. For ATM, CPU 1 seems to work harder and the difference between CPU 0 and CPU 1 is smaller. Again, some further testing over ATM would be required.

D. 2 to 8 requestors, 2 CPUs, compression

• From the figures obtained for this test set, it seems that adding more and more requestorsprocessors not running on the same CPU at the same time as compression causes more confusion than anything else.
• The performance edge of Fast Ethernet over Ethernet is much slower than it was in the previous test cases. It even scores lower in the 6 and 8 requestor scenarios.
• With ATM, although we do not have complete figures, we get throughput rates even lower than with Ethernet. Additional ATM testing would be required to sort out the interactions between components.
• The following table gives the approximate CPU consumption for the various transfers:

# ProcessReqs. CPU 0 CPU 1 CPU 0 CPU 1 CPU 0 CPU 1 CPU 0 CPU 1 K-1K-Eth K-1K-Eth S-7K-Eth S-7K-Eth S-7K-ATM S-7K-ATM S-7K-FastE S-7K-FastE 2 50% 30% 40% 40% 52% 40% 4 65% 40% 65% 80% 55% 45% 6 70% 45% 50% 45% 8 65% 40% 50% 40%

CPU consumption with ATM is higher than it is with other (Ethernet and Fast Ethernet) medium;