Test Setup
| Test System | |
|---|---|
| Processor | 2x Intel Pentium 4 Xeon, 2.8 GHz |
| Motherboard | Supermicro X5DL8ServerWorks GC-LE |
| RAM | 2x 512 MB PC2100/DDR266Registered, ECC - Samsung |
| IDE Controller | SB7440 UltraATA/100 Controller |
| System Hard Drive | IBM Deskstar 60 GXP, IC35040, 40 GB, 7,200 rpm |
| Test Hard Drive | 6x Seagate Barracuda 10K.4 (ST 3146807LW), 147 GB, 10,000 rpm, Ultra320 SCSI |
| Display Adapter | nVIDIA GeForce4 MX440, PCI, 64 MB |
| Network Card | Broadcom BCM5703, 1 Gb/s, on board |
| Operating System | Windows 2000 Server SP3 |
| Benchmarks and Tests | |
| Access Time | ZD WinBench 99 2.0 |
| Performance Tests | Intel IOMeter 2003.02.15 |
| I/O Performance | Intel IOMeter 2003.02.15 |
| Drivers and Settings | |
| Graphics Driver | nVIDIA 44.03 |
| Drivers | Windows 2000 Server default Chipset drivers |
| DirectX Version | 9.0 |
Because we have not used Intel's IOMeter software often up to this point, we would like to briefly explain the test procedure.
IOMeter allows access patterns to be generated for one or several drives, so that a typical system requirement can be re-created. For this article, we used four test processes:
- Maximum data transfer rate in MB/s
- AggregateBlock sizeProportion of read accessesProportion of random accesses100%64 kB100%0% (sequential)
No doubt this test is very theoretical and only serves the purpose of determining the maximum transfer rate. We are dispensing with write accesses and random accesses - only sequential reading is in demand.
- Maximum I/O performance in I/Os per second
- AggregateBlock sizeProportion of read accessesProportion of random accesses100%512 Byte100%0% (sequential)
In this area, we measured with 1, 4, 8, 16, 64 and 256 I/Os per drive access, since this can vary depending on the controller. Once again, this test only represents a theoretical result, as block sizes of 512 Bytes would hardly be used in practice.
- File server access pattern (recommended by Intel)
- AggregateBlock sizeProportion of read accessesProportion of random accesses10%512 Byte80%100%5%1 kB80%100%5%2 kB80%100%60%4 kB80%100%2%8 kB80%100%4%16 kB80%100%4%32 kB80%100%10%64 kB80%100%
In this respect, we measured with 4, 16, 64 and 256 I/Os for each access. The distribution of the required block sizes corresponds to the recommendation by Intel and simulates the environment of a typical file server quite efficiently. We used a total of six hard drives, which ensures high usage of the controller.
- Database access pattern
- AggregateBlock sizeProportion of read accessesProportion of random accesses100%8 kB67%100%
The last access pattern is used by other online magazines as well: with a block size of 8 kB, it reflects other database applications.
