The following discusses actual test results obtained whilst performing checks in order to decide to change network interfaces to 100Mb data rates.
The network in question is a small, single segment network supporting 10 users. The configuration and type of the equipment used can be viewed in appendix 1. The majority of users were using 'office' type applications with the data and executables located on the network server. The traffic was therefore in 'bursts' with few collisions.
In order to perform the check, two test files were created...
These file sizes were chosen as the 1Mbyte file is guaranteed to fit within the cache on the fileserver, the 100Mbyte file is guaranteed to exceed the cache.
Non-cached Test For this test, in a DOS window, a 100Mbyte file was copied...
a) ... from the server to the local disc (100Mbyte transfer)
b) ... from the server, back to the server via the client machine (200Mbyte transfer).
Cached Test
For this test, in a DOS window, the same 1Mbyte file was copied 100 times...
a) ... from the server to the local disc (100Mbyte transfer)
b) ... from the server, back to the server via the client machine (200Mbyte transfer).
A DOS batch file was used to automate the process (see note at the bottom of this document).
The results of the tests are shown in the table below;
| 10Mb Ethernet | 100Mb Ethernet | ||
Time | Data Rate | Time | Data Rate | |
Non-cached (100Mb files) | 120 seconds (100Mbyte) 240 seconds (200Mbyte) | 0.83Mbyte/sec | 59 Seconds (100Mbyte) 134 seconds (200Mbyte) | 1.7Mbyte/sec |
Cached (1Mb files) | 110 seconds (100Mbyte)
| 0.9Mbyte/sec | 17 seconds (100Mbyte) 30 seconds (200Mbyte) | 6Mbyte/sec |
Analysis of Results
With the 10Mbit ethernet, caching on the server does not appear to significantly alter the time of transfers, therefore it is likely that the bottle neck is the network media (cable/adaptor).
For this network, changing to 100Mb ethernet significantly speeds the operation of the system. The data rate is at least twice that of the 10Mb cable.
If the users were expecting a ten-fold increase (as may be expected moving from 10Mb to 100Mb media) then they will be disappointed. The reason is that the network is only as fast as its slowest link. With this set-up it is likely that the bottle neck is now in the server to disc channel. This could easily by enhanced by switching to Ultra SCSI, Wide SCSI (20Mbyte/sec) or Ultra-Wide SCSI (40Mbytes/sec) and suitable controllers.
The cached rate of 6Mbyte/sec is unlikely to be significantly improved upon. If a small number of large files are frequently accessed, then this caching could be extended by adding more memory to the server.
Note 1: The cache test ignores the small overhead of opening and closing the file 100 times, and concerns its self only with the overall time for the transfer. For this reason, the test results will return a slightly lower throughput than is actually occurring. This additional time is constant, regardless of the data rate on the cable, so the comparative timings are still valid.
Note 2: CPU utilisation of the server was low during all tests.