General overhead

In this subsection, we evaluate system overhead just from RightSpeed running unused in the background. There are two main sources of this overhead: (a) making the timer interrupt every 1 ms instead of every 10 ms causes interrupt-processing time to increase; and (b) filtering I/O requests to count them increases the time to perform each I/O.

To evaluate these effects, we ran the following benchmarks on a system with a 450 MHz Pentium III:

1. Read an uncached 32 KB file
2. Write a 100 KB file with write-through
3. Read 32 KB directly from the disk
4. Compile the RightSpeed logger device with the Windows DDK
5. Format a Ph.D. dissertation with L^ATEX
6. Perform a CPU-intensive mathematical loop

We ran them without any RightSpeed modules loaded, with only the RSIoCnt module loaded, with only the RSTask module loaded, and with both of those two modules loaded. In all cases, we disabled the network to avoid interference from network interrupts. None of these benchmarks use RightSpeed at all; indeed, we did not even install RSLib to perform these experiments. We ran each benchmark enough times that the 95% confidence interval about the sample mean included no values more than 0.01% away from the sample mean, or 10,000 runs occurred, or 2,000 seconds passed, whichever came first. Figure 3 shows results.

Figure 3: Time to perform various benchmarks without RightSpeed and with various components of RightSpeed enabled, shown with 95% confidence intervals. Note that the Y-axis origin is not zero.

We see that RSIoCnt adds 0.3-1.5% overhead, with an average of 0.5%, due to filtering I/O operations. If we did not have to use a file system filter to do this, e.g., if Microsoft provided hooks allowing one to simply count ongoing I/O's and be notified when the last I/O leaves the system, this overhead would likely be lower. We also observe that RSTask, by virtue of it reducing timer granularity from 10 ms to 1 ms, increases operation times by 0.7-1.6% with an average of 1.1%, presumably due to the system responding to more frequent timer interrupts. Combined, the overhead is 1.2% on average.