Understanding a Context Switching Benchmark
There is probably no way to make the search loop run faster; it is well-written code. However, it could be eliminated by simply keeping contexts in the ready queue in sorted order; then the next context to run is the context at the head of the ready queue. Keeping the ready queue in sorted order would require code to take each context being added to the ready queue and place it in the appropriate position. The time needed to find this position would add complexity to the scheduler. For large load averages (implying many contexts on the ready queue), there might be a considerable time savings, but in the more normal case of small ready queues, there is no significant savings.
Minimizing the time needed by the recalc loop would be easy. Again, it is well-written code not likely to be improved upon, but it need not run as often as it does. By changing sched_yield, the recalc loop can run much less often.
For Linux 2.0.30, sched_yield acts as if the yielding context has used the entire time slice. Instead, what if sched_yield acted as if the yielding context had used only one tick of its time slice? Several effects would be noticed:
The yielding process would have its priority reduced by one, rather than temporarily set to zero.
The recalc loop would run much less often. Generally, it would run 1/20 as often (depending on process priority).
The new sched_yield is shown in Listing 3. Compare it to the one in Listing 2. Only one more line is included, yet a large increase in performance is shown for this benchmark.
Table 2 summarizes the performance after this change was made. Note that as the run-queue length increases, both the NT and the Linux scheduler take longer to context switch. Linux starts out being the faster context switcher, but NT does relatively better as run-queue length increases. For run-queue lengths of 20 or less (almost always the case in real life), Linux is better.
By making a two-line change to the source code, this benchmark can be greatly improved. However, the benchmark arguably does not reflect real-life usage. Nevertheless, only a two-line change to the kernel is required for a significant benefit to a small number of users. After this change, Linux outperforms Windows NT in all aspects of process and thread creation and in context switching.
All listings referred to in this article are available by anonymous download in the file ftp.linuxjournal.com/pub/lj/listings/issue57/2941.tgz.
- Let's Go to Mars with Martian Lander
- Applied Expert Systems, Inc.'s CleverView for TCP/IP on Linux
- My Childhood in a Cigar Box
- Papa's Got a Brand New NAS
- Rogue Wave Software's TotalView for HPC and CodeDynamics
- Panther MPC, Inc.'s Panther Alpha
- NethServer: Linux without All That Linux Stuff
- Debugging Democracy
- Simplenote, Simply Awesome!