NetWare Internal and External Bridge Performance Benchmarking Drew F. Jackman Associate Consultant Systems Engineering Division Abstract: One of the key factors of network optimization is performance. Bridges are used on the local area networks to increase the performance of a particular section within the network. In doing this, the traffic between sections (over the bridge) suffers. A network administrator must decide which bridging configuration will give the best performance: internal or external, dedicated or nondedicated. When integrating a bridging configuration, all costs must be weighed against the advantages and disadvantages. Introduction This document describes two network bridging configurations for NetWare LANs (internal and external) and summarizes the benchmarking tests performed on these configurations. The benchmarking tests were designed to isolate the bridging machine as the device under test, with all other network elements kept constant. This method shows any change to the bridging machine reflected in the results of the performance tests (in the throughput). One test was conducted as a control, and is the standard to which all the other tests will be compared. Throughout this document the terms bridge and router are interchangeable, since the NetWare products operate at the network layer (third) of the Open Systems Interconnection (OSI) model instead of at the data-link layer (second). But in keeping with terminology originally used by Novell for this particular function, the term bridge is used. All 286 NetWare bridges can connect up to four subnets to make an internetwork appear as one logical network. Each bridge in a network keeps a dynamic routing table containing the addresses of the other nodes on the network. It contains the fastest route to each node, and for better reliability, it also contains a list of alternate routes that can be used if the primary route were to go down. Interesting points illustrated by this testing are: (1) the comparison between internal and external bridges, (2) the comparison between SFT NetWare v2.15 and NetWare 386 v3.0, (3) the degradation curve caused by adding external bridges in series, and (4) the performance comparison between a dedicated and a nondedicated bridge. Bridges NetWare bridges are internal and external. An internal bridge is contained within a file server and operates along with the normal file server functions. Thus an internal bridge is actually bridging information through a file server to another network or backbone on the LAN. An external bridge is a separate machine running the bridging software that connects multiple networks. This type of bridge can function in one of two ways: solely as a bridge (dedicated), or as a bridge that simultaneously functions as a workstation (nondedicated). A bridge functions as the buffer between two similar or dissimilar protocols. It can also be used to break one large network into two smaller ones to increase the performance of each by reducing the traffic on the wire. An external bridge used strictly as a bridge is called a dedicated bridge. A dedicated bridge does nothing but interconnect multiple networks by conveying data among them. An external bridge that functions as a bridge and as a workstation simultaneously is called a nondedicated bridge. In a nondedicated bridge the bridging software runs on top of DOS and the workstation shell. In addition to the possible degradation in processing speed of the workstation and bridge (due to both processes occurring simultaneously on the same machine), there is also a possibility that the other applications running on that workstation will malfunction and cause the bridging process to stop. All the workstations connected through the bridge/workstation will be cut off from the network on the opposite side of the bridge. Test description Each of the tests was configured and performed with the purpose of isolating the bridging section as the test module. Hardware The benchmark tests were performed with the following hardware configurations. Each of the tests was performed by accessing a Novell 286B file server set at 8MHz with 0 wait states. This file server was installed with a Novell Disk Coprocessor Board (DCB), a CDC WREN III embedded SCSI drive, and an NE2000 Ethernet adapter. The file server was accessed through the different bridging devices (see Figure XX). The dedicated and nondedicated internal 2.15 and 3.0 file server tests were performed using a Compaq 386 25MHz computer, installed with an ISADISK disk driver and 2MB of memory (the minimum amount needed to run NetWare 386). The nondedicated file servers were kept busy with a single Novell 286A workstation logged in and running the performance evaluation test program (PERFORM2 described below). The two internal configurations were run a second time using a 286B 8MHz computer installed with a Novell DCB. This enabled a comparison of the internal and external bridges. Each of the internal bridge tests used five Samsung 386S 20MHz computers and one Novell 386AE 16MHz computer as the workstations (see Appendix A). All internal bridge tests were performed using NE2000 Ethernet adapters in the file server, the bridging machine and the workstations. Thin cable Ethernet (RG-58 A/U) was used to connect the network. Figure 1: General block diagram of the bridging test setup The external bridge tests used the same six workstations used in the internal bridge tests. The computers used as the dedicated external bridges were 8MHz 286As, set at 0 wait states. The nondedicated bridge tests also used one (or two) of the same 286As as the bridging devices. All of the external configurations were connected by NE2000 Ethernet adapters. The only exception occurred during the multiple dedicated bridge tests where NE1000 Ethernet adapters were used in the workstations, with NE2000s in the bridges and the file server. During the read overlaid performance test (described below) the workstation adapter had little or no effect on the degradation of the final throughput because all six workstations were transmitting through one conduit-the bridge adapter and file server adapter. The workstation adapters wait for a turn to transmit through the bridge and file server adapters. Thus the NE1000s used in the workstations during the external dedicated tests had no effect on the final system throughput when compared with the NE2000s used in the workstations. To isolate the bridging device as the test module, the file server adapter and bridging adapter were not changed, but the bridging configuration was changed. Thus each test examined the bridging device only. Software The operating system on the file server of Figure XX was SFT NetWare v2.15. The bridging file servers (internal bridges) were booted with either SFT NetWare v2.15 or NetWare 386 v3.0. The external bridges were booted from the bridging software included in the NetWare 2.15 software package. The dedicated bridges were configured in real mode and the nondedicated in protected mode (see SFT/Advanced Bridges manual for more information). The tests were executed using the PERFORM2 (version 2.3) performance evaluation program,which has a record size of 4,096 bytes and 1,000 iterations for each test. Only the read overlaid function of PERFORM2 was used during the benchmarking tests, which enabled testing of only the bridging component. Workstations making overlaid read requests access the same data repeatedly. After the first workstation accesses the information from the hard disk, it is in the file server's cache memory to be used by all the workstations requesting it, eliminating the hard disk access speed from the test. Since all test configurations access the data from memory at the same speed, the file server memory access is also eliminated from the tests. This leaves only the degradation caused by the bridging device-the intended result of the tests. The NetWare 386 v3.0 internal bridge tests were run several times to allow the dynamic memory allocation to stabilize (see Technical Overview; NetWare 386, ppg. 21 and 22). Configuration The internal bridge configurations tested were as follows: Dedicated bridging v2.15 file server Nondedicated bridging v2.15 file server (nondedicated, in this case, meaning that it is busy performing its file server functions while it is bridging) Dedicated 386 v3.0 file server Nondedicated 386 v3.0 file server (nondedicated meaning that it is busy being a file server) The external bridge configurations tested included the following: Dedicated bridge configurations with from one to six bridging machines in series Nondedicated bridge Two nondedicated bridges in series Test Results The first configuration tested was the control model with no bridging. The resulting throughput was 623 Kbyte/s. This value was used to compare the degradation experienced by the other bridging configurations. Internal The results of the internal bridge tests demonstrate the immediate performance degradation caused by bridging. As displayed in Figure XX, a substantial performance drop occurs when network packets or requests pass through a bridging device. For this configuration, the drop is a 41 percent decrease in throughput. This is the best possible throughput for this configuration. The next point illustrated by the internal bridge tests is the performance comparison between the two versions of NetWare (SFT v2.15 and 386 v3.0). As shown in Figure XX both types of file servers functioned at approximately the same level of throughput. Figure 2: Internal bridge comparison External A NetWare external bridge can be configured as dedicated or nondedicated. The first test performed on the external bridges involved adding from one to six dedicated bridges in series and testing the throughput of the system as each bridge was added (see Appendix A). The results of this test show the resulting performance degradation curve (see Figure XX). This happens when information passes through an internet with multiple bridges between the accessing workstation and the intended file server. A similar performance curve would occur when information passes through multiple internal bridges. Figure 3: Multiple bridge degradation curve The initial degradation through the first bridge results in a 43 percent drop in throughput. Each additional bridge adds a small percentage to this initial performance degradation. Table I contains each percentage, from one to six bridges, as compared to the configuration with no bridge (no bridge is equal to a zero percent drop in throughput). For example, for six bridges in series, there is an 81 percent throughput degradation compared to the no bridge test. This makes the throughput of a six bridge network only 19 percent of a no bridge network's throughput. If more bridges were added, the curve in Figure XX would continue in a steady decline. After a total of sixteen bridges, NetWare would eliminate the packet to ensure that the packet would not continue through a circular internet connection. Table I: Percent performance degradation compared to no bridge Number of bridges 1 2 3 4 5 6 Performance drop 43% 52% 60% 72% 76% 81% The nondedicated bridge tests were run by operating the performance evaluation test (PERFORM2) in two network drives (see Figure XX). One directory had six workstations simultaneously performing the test, while the bridges/workstations were in another directory running the same type of test. The workstations were physically connected through the bridge(s). Figure 4: Nondedicated bridge hardware and software setup The results show that the nondedicated bridge is the slowest of the bridging configurations. With one nondedicated bridge the throughput drops 60 percent. With two nondedicated bridges in series the performance drop is 67 percent. Figure XX illustrates the performance of the nondedicated configuration as compared to the test using no bridge. These results represent the throughput for the six workstations, not the throughput of the bridge/workstations. During the nondedicated test, the bridge/workstations were cut to less than 10 percent of the throughput shown in the test using no bridge. The workstation installed with the bridging software was slowed down to approximately 30 to 60 Kbyte/s. Figure 5: Nondedicated bridge degradation A performance comparison between the two types of external bridges is shown in Figure XX. Because the nondedicated bridges are functioning simultaneously as workstations, throughput is slowed. A user can expect an average of 30 percent less performance (throughput) through a nondedicated bridge. Figure 6: Comparison of dedicated and nondedicated bridge performance The next comparison that can be drawn from the bridging tests is between internal and external bridges (dedicated and nondedicated). The results of the internal and external bridge comparison show the internal bridge has a 10 percent slower throughput. If the nondedicated configurations are compared, the internal bridge has a 15 percent slower throughput (see Figure XX). If the internal bridge is heavily loaded with file server tasks, its throughput is even slower. Remember this set of internal bridge tests used a 286B file server so the comparison between internal and external would be valid. Figure 7: Comparison of internal and external bridges These results indicate the nondedicated internal bridge configuration is the slowest type of bridge when a file server with a slow processor is used (clock and processor). By picking a fast file server carefully, this problem can be eliminated. If the bridging file server is heavily loaded down, the throughput will decline whether the file server is fast or not. Conclusion The information contained in the test results shows some important points that need to be emphasized. The first is the unavoidable performance degradation (throughput degradation) that occurs when information passes through a bridging device. The second is the comparison between external and internal bridges. The third is throughput comparison between the two versions of NetWare (286 v2.15 and 386 v3.0), and the fourth and final point is the curve obtained as information passes through multiple bridges. When a bridge is placed on an internet, the best possible throughput for the information passing through the bridge is approximately 60 percent, as compared to the throughput of an internet without the bridge. If multiple bridges are added, or the bridging machine has a slow clock speed and processor, the throughput will be reduced even more. Figure 8: Comparison of performance drop corresponding to each type of bridge There are two things to consider when deciding between an internal and an external bridge. The first is the overall cost of adding the new software and hardware. The second is the performance (throughput) achieved with the chosen bridge type. For example, if there are multiple file servers on one backbone, and some of the workstations on that backbone need to access a second backbone, one option is using an internal bridge. The only additional cost is the installation of an extra adapter in one of the existing file servers, to make it the internal bridge between the two backbones. The performance would depend on how busy the new bridging file server is, and how fast the file server hardware operates (clock and processor speed). The throughput, compared to no bridge, would be from 40 percent to 60 percent. This would be a reliable bridging configuration (see Table II). The best solution in this case is a fast file server that is not heavily used. If a dedicated external bridge is chosen, the average performance would be better than with an internal bridge, but the cost is greater. The throughput would stay around 60 percent. A new machine with two new adapters is needed as the bridge. This would be the best choice if performance maximization and reliability are the ultimate goal. The final possible configuration in this situation is using an existing workstation as a nondedicated external bridge. The only expense is the extra adapter. This configuration is inexpensive, but would also result in a very slow throughput. It could also cause malfunctions between the two backbones periodically. The system throughput would average around 40 percent of the configuration using no bridge, and the bridge/workstation throughput would be less than 10 percent. Table II: Advantages and disadvantages of bridging configurations compared Option Reliability Performance Additional Restrictions No bridge High 100% None Cannot split traffic Internal Medium 40-60% One extra Bridging server can be tied up Dedicated Medium 60% Extra machine None external and two adapters Non- Low 30-40% One extra Applications on the dedicated adapter BR/WS can hang. Performance across the BR/WS about 10 percent. The reliability factor is a gauge of whether the bridge is likely to go down. Another question to address when creating an internal bridge is which operating system will give the best throughput, 286 v2.15 or 386 v3.0. The answer is simple-both operate the bridging function equally well. Use whichever is most convenient. Whether they will perform equally when both file servers are heavily loaded remains to be seen in a future test. With a large internet information may pass through more than one bridge. As shown in Table II, the initial drop through the first bridge is about 40 percent (making the throughput 60 percent as compared with no bridge). Each additional bridge drops the resulting throughput an average of eight percent. If throughput from one backbone to another is vital, then a one (or a no) bridge maximum should be designed into the internet. The criteria for deciding which configuration of bridges to use in an internet follows: No bridge: high throughput unless there is high traffic, no extra cost External dedicated bridge: consistently high throughput, high setup cost External nondedicated bridge: low throughput, medium to low setup cost Internal bridge: medium to high throughput, low setup cost Appendix A Figure 9: Test setup for dedicated internal bridges Figure 10: Test setup for nondedicated internal bridges Figure 11: Test setup for multiple external bridges Appendix B Table III: Data gathered from the benchmarking tests Configuration Throughput Utilization Utilization Kbyte/s bridging accessed file server file server No bridge between the file server and workstations 622.92 N/A 98% Internal bridge in a NetWare v2.15 363.57 69% 97% dedicated file server using 386 processor Internal bridge in a NetWare v2.15 366.88 87% 97% nondedicated file server using 386 processor Internal bridge in a NetWare v3.0 365.20 68% 97% dedicated file server Internal bridge in a NetWare v3.0 364.26 85% 95% nondedicated file server Internal bridge in a NetWare v2.15 320.61 100% 85% dedicated file server using 286 processor Internal bridge in a NetWare v2.15 213.37 99% 56% nondedicated file server using 286 processor External dedicated bridge 365.39 N/A 93% Two external dedicated bridges 295.97 N/A 78% Three external dedicated bridges 251.86 N/A 67% Four external dedicated bridges 172.52 N/A 45% Five external dedicated bridges 150.83 N/A 40% Six external dedicated bridges 119.50 N/A 30% External nondedicted bridge 250.31 N/A 65% Two external nondedicated bridges 206.72 N/A 57%