Tuning LM on a Token-Ring Network (73283)

MORE INFORMATION

Token Ring performance depends heavily on the type of machine/bus/net card used for the measurements. Using a PS/2-IBMtok (MCA) as a server increases the total throughput by 56 percent over that of a Compaq386/25-IBMtok (ISA). Other studies on this subject indicate that the total throughput of the server can increase as much as 9.6 times by using Advance Micro Channel (AMC), instead of ISA. Therefore, given a choice, one should use higher bandwidth, DMA bus master card, and choose better bus architecture, such as AMC or ISA. These results should be verified in-house.

[Sniffer data shows that back-to-back maximum frame (2000 bytes) transmission is 3 milliseconds. This translates into the total maximum throughput of 651 KBPS (kilobytes per second). Considering other delays involved in transmitting data, this number justifies the maximum effective throughput obtained (536) for all the six workstations. Sniffer's timer granularity of 1 millisecond could have also contributed to the error.]

Shared Memory

The IBM Token Ring card has 64K RAM on board. The card can be configured to use 8K, 16K (with RAM paging 64K), 32K, and 64K of RAM. It is better to use a larger size of RAM because this allows the user to specify a larger buffer (xmit/recv) sizes. It was discovered during these measurements that whenever there is less memory available than the total buffer size specified:

• No error is indicated during driver load/bind time.
• "Net3100 The operation failed because of a network software error occurred" is displayed while doing any network activity.

Transmit Buffer Size

This parameter dictates the frame size used. Token Ring allows a maximum frame size of 4K for 4 MBPS (megabytes per second) and 17952 bytes for 16 MBPS. The frame size used on a session is equal to the minimum of the transmit buffer sizes specified (or used) for the server and the client. Default transmit buffer size is a minimum of 25 percent of RAM or the maximum size allowed.

An interesting fact is that NetBEUI currently (LM2.0C) uses the maximum frame size of 2000 bytes. Therefore, specifying xmitbufsize greater than 2K results in memory wastage. In the future, this limit will be changed to 18000 bytes.

Changing the xmitbuffer from 256 bytes to 1K increases the throughput by 36-153 percent; changing it from 1K to 2K results in a 16 percent throughput increase. Changing the xmit buffer size beyond this value does not increase the throughput. At higher transmit buffer sizes, write (from clients to server) performance degrades because less memory is available for receive buffers. This degradation occurs only when less RAM is specified/used.

Number of Transmit Buffers

Increasing this value from the default of 1 to the maximum value (2) increases the server throughput by 13-18 percent for large I/O. Therefore, it's better to use two xmit buffers if there is enough memory onboard.

Receive Buffer Size

Changing this from 256 bytes, 1K, to 2K did not make any significant difference in the total throughput. The write performance for large sizes degrades 8-9 percent for larger receive buffer sizes on workstations. For single workstations, higher values of receive buffer size (2000) improves the read performance by 7-9 percent. Specifying a recvbufsize value higher than 2000 bytes results in a network software error while doing any network activity (DCR).

Number of Receive Buffers

It is OK to leave this value at the default (2) because all the memory available is used for receive buffers.

Early Release

This parameter is supposed to increase the overall throughput by releasing the token after transmitting a data frame. Performance tests conducted with six workstations did not show much difference between early releases and without it. However, this will definitely become a significant factor in cases where the number of machines on a network is large (causing larger token/frame forwarding overhead and also longer propagation delay). Even with this setup, early releases write performance can degrade because a large number of back-to-back frames may cause the server to drop some frames. This problem may not occur for higher RAM (64K) and faster net cards (MCA, busmaster).

Machines in Test

Servers: Wkstas: