4.1.6 I/O Performance using IOstone Benchmark

To compare the I/O performance of our device over the Ethernet with the traditional magnetic disk of the same bandwidth we ran the IOstone Benchmark.

IOstone is meant for benchmarking file I/O and buffer cache efficiencies. It does this by creating NSETS (4) of SET_SIZE (99) files. Then iostone performs I/O on each file in each set. The type of I/O is randomly picked (read or write).

More detailed description of the benchmark process is :

• Allocates space for the filenames of NSETS (4) of SET_SIZE (99) files each, and creates the filenames.
• Reads in some files to flush the buffer cache.
• Order filenames in a random permutation for reading/writing.
• Start the timer (using ftime() ).
• Perform reads/writes (read:write = 2:1) to the files ITER (4) times.
• Stop the timer and prints the iostones of the system ( a constant divided by the number of msec that were measured ).

 
Table 4:   IOstone Benchmark performance

The comparison of IOstones/second can be seen in Table 4. It is obvious from the table that the Network RamDisk even over the Ethernet, and using TCP/IP has over seven times higher performance, than the magnetic disk.

 

 

Figure 11: Comparison of the I/O Latency performance of the Linux NRD Client on the Magnetic Disk vs. Network RamDisk over the Ethernet and the ATM interconnection network, using HBench-OS: This test measures the latency of performing certain file system metadata operations, in particular file creates.The latency shown is the number of milliseconds (ms) per file creation for a variable number of 2KB file creations on each of the devices. The benchmark were executed from user level on a Pentium II 233Mhz PC with 64MB of main memory running Linux 2.0.33, and compiled with the GNU C compiler (gcc 2.7.2.1). The Network RamDisk used as servers two Sun Ultra 1 workstations, each with 128 MBytes of main memory and Sun ATM NIC, running Solaris 2.6 .