Check out the new USENIX Web site. next up previous
Next: Quantifying Service Inversion Up: Service Inversion Previous: Service Inversion

Identifying Service Inversion

To qualitatively understand the prevalence of service inversion, we take the latency CDFs from Figures 6 and 7 and split them by decile. Since SPECWeb biases toward small files and more than 95% of the requests could fit into physical memory, ideal response times would be roughly proportional to transfer sizes. By examining the different response sizes within each decile, we can estimate the extent of reordering. To simplify the visualization, we group the responses by sizes into four series such that their dynamic frequencies are roughly equal. The details of this categorization are shown in Table 3.

Table 3: Workload categories for latency breakdowns. Percentages shown are the dynamic request frequencies for the given file sizes in the SPECWeb99 workload.
series size range percentage  
1 0.1 - 0.5 KB 25.06%  
2 0.6 - 4 KB 28.05%  
3 5 - 6 KB 23.55%  
4 7 - 900KB 23.34%  

The graphs in Figures 10 and 11 show the composition of responses by decile for the two servers, with the leftmost bar corresponding to the fastest 10% of the responses and the rightmost representing the slowest 10%. These graphs are taken from the latency CDFs at a load level of 0.95.

In a perfect scenario with no service inversion, the first 2.5 bars would consist solely of responses in Series 1, followed by 2.5 bars from Series 2, etc. However, both graphs show responses from the different series spread across all deciles, suggesting both servers exhibit service inversion. One surprising aspect of these plots is that the Series 1 values are spread fairly evenly across all deciles, indicating that even the smallest files are often taking as long as some of the largest files.

Figure 10: Apache CDF breakdown by decile at load level 0.95

Figure 11: Flash CDF breakdown by decile at load level 0.95

Some inversion is to be expected from the characteristics of the workload itself, since directories are weighted according to a Zipf-1 distribution. With roughly 600 directories in our data set, the last directory receives 600 times fewer requests than the first. So, even though files 100KB or greater account for only 1% of the requests (35 times fewer than the smallest files), the directory bias causes the largest files in the first directory to be requested about 17 times as frequently as the smallest files in the final directory. While the large files still require much more space, an LRU-style replacement in the filesystem cache could cause these large files to be in memory more often. In practice, this effect is relatively minor, as we will show later in the paper.

next up previous
Next: Quantifying Service Inversion Up: Service Inversion Previous: Service Inversion
Yaoping Ruan