Check out the new USENIX Web site. next up previous
Next: Results Portability Up: Latency Previous: Diskbound static workload

Excess parallelism


We also observe that all servers tested show latency degradation when running with more processes, though the effect is much lower for our new server. This observation is in line with the self-interference between the helpers and the main Flash process which we described earlier. We increase the number of helper processes and measure its effect on the SpecWeb99 results, as shown in Table 9. We observe that too few helpers is insufficient to fully utilize the disk, and increasing their number initially helps performance. However, the blocking from self-interference increases, eventually decreasing performance. A similar phenomenon, stemming from the same problem, is also observed with Apache. Using DeBox, we find that Apache with 150 processes, sleeps 3667 times per second, increasing to 3994 times per second at 300 processes. This behavior is responsible for Apache's latency increase in Figure 11.


Table: Parallelism benefits and self-interference - The conformance measurement indicates how many requests meet SpecWeb99's quality-of-service requirement.
# of helpers 1 5 10 15
Blocking count 114 295 339 394
% Conforming 40.9% 95.1% 96.9% 89.5%



This result suggests that excess parallelism, where server designers use parallelism for convenience, may actually degrade performance noticeably. This observation may explain the latency behavior reported for Haboob (41).



next up previous
Next: Results Portability Up: Latency Previous: Diskbound static workload
Yaoping Ruan
2004-05-04