Powering Up the System
For my performance tests, I used the Lab's standard configuration: a set of client machines on a 100Mbps Ethernet network that simulates the workload of multiple users. (For details about the Lab's test environment, see the sidebar, "The Lab's Test Environment.") I used Bluecurve's Dynameasure/File 1.5 and Dynameasure/SQL 1.5 as my workload engines. (For information about the Dynameasure products, see "Dynameasure Enterprise 1.5," September 1997.) The Lab's test environment, with its Dynameasure software, provided quantitative benchmarks that I used to compare the different machines' performance.
To test the Northbridge NX801's file and print services performance, I compared the system with a brand-name SMP server that performed well in recent Lab tests. The brand-name server has four 200MHz Pentium Pro processors, 1GB of RAM, four 10/100Mbps Fast Ethernet PCI network cards, and four 4GB Seagate Barracuda 4LP SCSI hard disks. I configured the brand-name server with NT Server 4.0 Standard Edition and SP3.
I compared the two systems using Dynameasure/File's Copy All Bidirectional tests. These tests perform in random order 16 transactions that copy compressed data, uncompressed data, binary files, text files, and image files between the server and clients. The test files range in size from 500KB to 5MB. The test specifications called for six steps that started with 10 simulated users and increased the number of users at each step to a maximum of 100 simulated users at the sixth step.
Graph 1 shows the two systems' throughput, which measures their data transfer capacity. For example, a system that copies a 1MB block of data in 1 second has a throughput of 1MBps. Both systems reached their maximum throughput in step 4: The Northbridge NX801 reached 6.30MBps, and the brand-name server reached 5.47MBps. Graph 2 shows the systems' average response times, which measure how quickly the systems can read a file and copy the file to another disk at each step. In step 4, the two systems' speeds were relatively close--3.74 seconds for the Northbridge NX801 and 4.19 seconds for the brand-name server. The typical user at step 4 would experience a half-second difference in speed between the two servers. Graph 3 shows the systems' motors, a measure of how many users can execute Dynameasure transactions during a step without dropping out because of environment stress conditions. The two systems' number of virtual users remained close through all the test's steps.
The Northbridge NX801 outperformed the brand-name server in throughput, speed, and motors at almost every step. If you are looking for a scalability solution for your enterprise's file and print services, focus on step 5 in all three graphs. When I increased the number of virtual users simultaneously on the servers from 75 to 85, the Northbridge NX801's throughput was 5.81MBps, compared with the brand-name server's 3.05MBps. The Northbridge NX801 had an average response time of 7.94 seconds, compared with the brand-name server's 23.89 seconds. The Northbridge NX801 costs about $20,000 more than the brand-name server; nevertheless, the Northbridge NX801 might be cost-effective for your enterprise. For a system of 85 simultaneous users, the Axil machine offers nearly double the throughput and one third the average response time of its brand-name competitor.
Kicking the System into Afterburner
For my tests of the Northbridge NX801's SQL Server capabilities, I pitted the system against another machine, one of the Lab's SMP servers. The Lab's SMP server has performed admirably in SQL Server tests and has crushed most of its challengers. This popular brand-name server has four 200MHz Pentium Pro processors, a 512KB Level 2 cache, and 512MB of RAM. I configured both systems identically with SQL Server 6.5 Standard Edition.
Using Dynameasure/SQL, I placed a measured stress on both systems using OnLine Transaction Processing (OLTP) SQL workloads. The test specifications included a 500MB dataset on each of the servers and Dynameasure/SQL's Order Entry/Mixed Reads and Writes test. This test consists of 19 SQL transactions, which randomly perform reads and writes against a database. The transactions simulate orders in a business database, including item entries such as customer information, payment record, and product inventory. The transactions provide a well-balanced test of system performance without isolating any subsystem. I tested the servers in six steps, each with an increasing number of simulated users ranging from 100 to 600.
Graph 4 shows the servers' performance capacity in terms of their number of transactions per second. The Lab's SMP server peaked at 93.50 transactions per second at step 6--great performance compared with other four-processor systems. However, the Northbridge NX801 reached 112.93 transactions per second at step 6. Graph 5 shows the systems' average response time per step. At step 6, the Lab's SMP server had an average response time of 1.57 seconds. The Northbridge NX801 had an average response time of only 0.36 seconds. Graph 6 shows the systems' motors per step.
The Northbridge NX801 was fast, and its transactions-per-second rate showed no sign of slowing. So I increased the number of simulated users on both systems to 1000 and reduced the test's think time. (Think time is a parameter within Dynameasure/SQL that lets you adjust the length of time between the completion of one transaction and the start of the next transaction to simulate the time actual users would require to think about the results of one database query before beginning their next query.) My adjustments to the system caused an increase in CPU and disk stress that the Lab's SMP server couldn't take. Beyond 600 users, the SMP server's number of transactions per second declined, and its average response time climbed exponentially. The Lab's SMP server could not support more than 668 users.
In contrast, the Northbridge NX801's number of transactions per second continued to rise linearly as I increased the number of virtual users, and the system's average response time never rose above 0.4 seconds. The system supported a maximum of 858 users, but the Northbridge NX801 didn't limit the number of virtual network users. I couldn't test the system beyond 858 users because of the Lab's benchmarking network limitations: The control server Dynameasure uses to conduct tests and measure results created a bottleneck within the network. The control server couldn't keep up with the Northbridge NX801, and its slower speed caused motors to drop out of the test. Using the Lab's resources, I could not overstress the Northbridge NX801.
The Northbridge NX801 is the benchmark system for SQL Server tests. Now that Microsoft has released NT 4.0 Enterprise Edition and SQL Server 6.5, the Northbridge NX801 has enormous performance potential. If you're in the market for a SQL Server system that can handle large network capacity, the Northbridge NX801 might be a smart investment.
End of Article
