®MDNM¯Introduction PC Magazine LAN Benchmark README.TXT File ------------ The PC Magazine LAN benchmark program is used to evaluate LAN hardware and software. This file provides an overview of the test programs and instructions on how to use them. The following sections are included in this document: LAN Test Overview How to Run the Tests Cleaning Up and Filenames PC Labs Setup How We Interpret Results You can write or call PC Labs at (212) 503-5570 if you have comments, suggestions, or questions. LAN Test Overview ----------------- The PC Magazine LAN benchmark test provides a way to impose a consistent load on a LAN and to time specific procedures. The tests generate repeatable results, which can then be compared with other similarly configured LANs with the same number of workstations. The following captions are printed in PC Magazine when LAN test results are reported. They provide an overview of the three types of file loading used in the tests. The PC Labs LAN benchmark tests are written in C and are independent of commercial software. We ran the tests on a test- bed of five 8-MHz IBM PC ATs. For our test-bed to better simulate the conditions on a medium-size network of 20 or more workstations, we have designed these loading tests so that a single station represents five to ten times the load of a user performing an interactive task (for example, updating records) on a network. By themselves, the elapsed times reported in these tests are not meaningful. They are valuable only when used to compare the performances of two or more systems running under near-identical conditions. Accordingly, we include the tests run on our Editor's Choice configuration of a 3Com 3Server3, 3+Share software, and EtherLink interface cards to provide a point of comparison. We also show results from a network of Novell's Advanced NetWare/286, EtherLink cards, and an IBM PC AT as the server. Advanced NetWare is our Editor's Choice for networking software, and our tested configuration is a typical one. The times are in seconds. Network Speed Under Load Results Stations 3+Share Advanced Netware/286 -------- ------- -------------------- 0 306 264 1 423 280 2 529 301 3 651 310 4 761 322 Hard Disk Access Load Results Stations 3+Share Advanced Netware/286 -------- ------- -------------------- 0 155 136 1 227 150 2 330 162 3 419 174 4 522 182 Database Load Results Stations 3+Share Advanced Netware/286 -------- ------- -------------------- 0 155 136 1 298 169 2 425 212 3 585 280 4 669 305 The Network Speed Under Load and the Hard Disk Access Load benchmark tests measure the time needed to perform a standardized task on the network. While the actual work loads used for these two tests (described below) are different, we used the same procedure for both. To obtain the elapsed times shown here, we ran a benchmark program performing a sequential create, a sequential read, a sequential write, a random read, and a random write of a large file. The record sizes used in these activities systematically rotate between 16K, 4K, and 512 bytes. The numbers shown in the three-dimensional chart are the total time necessary for all of these operations. We ran the test on all our ATs to load the network while timing just one of them. We then reduced the number of workstations one at a time to show the effect of loading on the network. The Network Speed Under Load test puts a heavy load on the network interface (cards, media, and so forth) while placing a minimal load on the hard disk by having each station continuously read and write its own 1-byte data file, changing the data each time. For systems with disk caching, the load on the hard disk is even smaller, since cached systems typically perform a disk write but do not require a physical disk read. The Hard Disk Access Load test heavily loads the hard disk and disk-caching system. To do this, each station randomly accesses its own 100K data file using 1K records. Data written to the file is changed each time. The random reads typically access data outside the cache, which forces a disk read, as does any write. The Database Load test exercises the system's record-locking support and the way it handles a number of random simultaneous accesses to a common file. This test times how fast each loading station accesses a common database consisting of an index and a data file. Half the accesses are simple searches of the index file and an accompanying access to a record in the data file. One quarter of the accesses perform the same operation but also lock the data record and update its contents. The remaining accesses update the index file and a data record. The index file is locked during every update and the DOS 3.1 RLOCK statement prevents simultaneous index file updates. How to Run the Tests -------------------- The LAN benchmark program uses only DOS 3.x file handling facilities and does not require LAN-specific support such as NETBIOS support. The test configuration requires at least one file server, which can be dedicated or nondedicated. The network must also have one or more DOS workstations attached. The network being tested should be set up and running according to the vendor's specifications before the benchmark program is run. Each workstation involved in a test should be logged onto the network, and the LAN benchmark program must be running on each workstation. Any network options related to file sharing that may be necessary for running the benchmark program should be implemented before the tests are run. The DOS SHARE program is often required for the database load procedure. Also, some systems require special attributes to be set for these files. Check the corresponding network documentation on file sharing for details. The LAN benchmark program is menu driven but it does require some manual coordination and setup. These options are chosen using the setup menu, which creates a setup file--PCMAGLAN.ARG. The LAN benchmark setup program should be run first. It is available from the main menu of the LAN benchmark program. You can choose from among various testing parameters--such as the file sizes to be used in a test. The setup file should then be saved. A single setup file can be shared if the file is available on a shared directory on the LAN or individual copies can be used on each workstation. Select a LAN load procedure on each workstation attached to the network. These workstations access data on the file server attached to the network--and are the cause of the contention for processing power. Only one workstation runs a time procedure to obtain the final numerical results. Remember to start the load procedures before starting the time procedure. The timed test runs to completion and should not be interrupted. It can be terminated using the Control-Break keys, but no results will be saved and the test must be restarted completely. The load procedures can be incrementally started or stopped while the timed test is running without having to restart each load procedure from scratch. It is more likely that you will manually terminate the load procedures after each portion of the time test is completed. The load procedures are interrupted using the Esc (escape) key and a prompt will allow the procedure to be terminated or to continue. The load procedures are options 1, 2, and 3 on the main menu. Load procedures 1 and 2 use individual and unique files for each workstation, and the files can be contained in different directories on a shared hard disk. The database load procedure uses a common set of files for all load workstations. The files used by the load procedures must be contained on the same network file server; otherwise, the load procedure may have little or no effect on the timed results. Load procedures 1 and 2 and the timed test automatically create their work files and delete them before the procedures terminate. Load 3, the database load, requires predefined data files that are built using the B option on the main menu. The size of these data files is specified in the setup menu. The size of the file is specified in terms of levels where smaller numbers equal smaller files. The sizes grow exponentially with respect to the number of levels. The timed test is option T on the main menu. It performs standard file operations on a single file, including creating the file as well as sequential and random reads and writes. The file and record sizes can be specified in the setup menu and the numbers can have a trailing K or M for kilobytes or megabytes, respectively. Record sizes can be in bytes, although the file size is rounded up to the nearest kilobyte. Up to eight record sizes can be listed and the record sizes are processed in the order listed. The test file is deleted after each record size is used and a new file is created for each additional record size. The program provides two result files. One is suitable for printing or incorporation into a document using most word processors (PCMAGLAN.TXT). The second is formatted for spreadsheet programs that can read comma-separated variable (CSV) format files (PCMAGLAN.CSV). The numbers saved in the timed test result files are identical; only the format is different. The LAN benchmark program will append new results to the end of the files if they already exist. Cleaning Up and Filenames ------------------------- Normally the timed procedure and loading procedures delete any files they create when the test is done. The exceptions are the files created for the database load procedure. These files must be deleted manually using the DOS DELETE or ERASE commands. The names of the work files used by the LAN benchmark program can be changed in the setup menu, but the defaults include: PCMAGLAN.IDX database index file PCMAGLAN.DAT database data file TMPXXXXX temporary work file The temporary work-file names actually have the XXXXX replaced by a random number so that each file is unique. The temporary work files may be accidentally left on the disk if the network or program is terminated improperly or prematurely aborted. Temporary files will never be used again, so they can be deleted at any time. Although the database files can be easily reconstructed using the LAN benchmark program, they should be deleted only when they are no longer needed. PCMAGLAN.ARG is a file created from the setup menu. It contains the user-defined options. A single copy of PCMAGLAN.ARG can be placed on a shared directory on the file server or each workstation can have its own copy. PCMAGLAN.ARG is loaded by default when the program is started. There is no way to change the default filename. However, if you want to rename the file in order to save different setups, you can load it explicitly through an option found in the main menu. Different setup files can be used to save different test parameters, different LAN names, and different result-file names. The setup file PCMAGLAN.ARG does not have to exist for the LAN benchmark program to run. Default values are used if there is no file. PCMAGLAN.TXT and PCMAGLAN.CSV are the default names for the result files created by the time test. These can be deleted if the results are no longer needed or can be saved accordingly. All filenames can contain drive designations and directory path names. This allows more flexible placement of the files and allows results to be placed on a drive other than the one being tested. The LAN benchmark program traps some errors, but occasionally "disk full" errors or LAN-related errors occur. It is also possible to corrupt the LAN benchmark database files. The program will detect and report on the latter. It is best to delete all temporary and database files and regenerate the database if this occurs. The results for any timed test should be disregarded if these types of errors occur. In general, the LAN benchmark program can be aborted using Control-Break. Reboot or power-down the workstation if the program must be terminated and it does not respond to Control- Break. Be careful when editing and appending time test results. Some text editors place an end-of-file character at the end of a file. This character is not visible when the file is typed or edited. The time test appends data to the physical end-of-file, which may appear after this character if the result files are edited and saved. The data will be in the file but will not be accessible using this type of editor or via the DOS TYPE command. Edit only copies of files from the time test to prevent this problem from occurring. Also, look at the file size versus the amount of information that is viewable to see if this is a problem you may have encountered. Cummulative time test results are appended to the end of a result file, not the beginning. This is a useful way of keeping the results for different numbers of workstations and different kinds of loads. PC Labs Setup ------------- The results printed by PC Magazine are obtained by repeating the LAN benchmark test using different load procedures and different numbers of load workstations. PC Labs uses standard IBM PC AT computers running at 8 MHz. Six ATs are used as workstations and an additional AT is used if a dedicated AT server is required. The dedicated AT server is augmented with 2MB of extended memory when it is used. The other workstations have no extended memory. One AT is designated as the timed workstation while the others are load workstations. The PC Labs configuration is set up in a single room so that each workstation is easily viewed at the same time. When PC Labs tests a non-IBM server, the vendor's server replaces the AT. Servers are tested as delivered. The LAN is set up based upon the vendor's documentation. No tuning is done on the server or the workstations other than what is specified in the standard installation procedure. All configurations are tested using supplied service programs and normal DOS applications to make sure the LAN is running properly before the LAN benchmark program is run. How We Interpret Results ------------------------ As described above, the benchmark-test results reported in PC Magazine are generated on a standard network configuration. Although you can test any LAN using the PC Labs benchmark program, results obtained using a LAN configuration different from the PC Labs standard configuration cannot be compared with results printed in the magazine. The LAN benchmark program tests file-access performance only. It does not test other aspects of the network such as print spooling, electronic mail, communications, or bridges. The three different load tests come into play as more workstations log onto the network. These tests are designed to place an unusually heavy and consistent load on the network. Bear in mind that the number of workstations a network configuration can support is normally much larger than the standard PC Labs LAN test configuration. The Network Speed Under Load test is designed to place a heavy load on the network software and transport mechanism. Each workstation reads and writes a single-byte file. This should generate a minimum amount of disk traffic on the server. In fact, some servers with good buffering techniques will perform no disk accesses until the tests are done. On the other hand, the test should generate a great deal of network traffic. The network adapter on the server can be kept at maximum load with a sufficient number of workstations. The Hard Disk Access Load test also generates a good deal of network traffic, but it also performs a large number of disk operations. The size of the data files used can be specified and should be larger than the amount of buffering available on the hard disk. The traffic is similar to that encountered when copying large amounts of data or when running applications that heavily access the file server's hard disk. The Database Load test uses a common database, which is accessed by each load workstation. Record locking is used on the index file and data file. Access to the records is random with a distribution of data record reads, data record updates, and index record updates. The load simulates a heavy access to a common database. This would be comparable to a significantly larger number of workstations performing query operations to a common database. The locking procedures are similar but not identical to many commercial network database products. However, the test is primarily designed to show how good or bad a network supports simultaneous file access with record locking. DOS 3.x file and record locking are mandatory for this test. The time test can be used in conjunction with any other load procedures you create. PC Labs tests only the same type of load within individual tests for consistency and to simplify interpretation of test results. The change in the test results shows how much variance may be observed on a heavily loaded or heavily populated network. The time test results can be interpreted in isolation based on the nature of the tasks (i.e., network traffic vs. server disk access). Results can be compared on the same network with varying numbers of workstations logged on (i.e., network contention for the same task with three workstations as compared with eight). Or you can compare a similar number of workstations using the same load but different network configurations (i.e., token-ring vs. star topology). For example, while write operations normally take longer than read operations, the factor will vary depending upon the network configuration. Likewise, you can isolate the effect of additional buffers on the server, as well as the workstation, by changing the setup options and rerunning the tests.