Performance under Different Workloads

In order to understand how the WOLF and LFS perform under different workloads, results for the four synthetic traces and four real-world traces are compared in Figure 4.

**Figure 4:** Overall Write-cost versus Disk Utilization under different workloads. *WOLF with 2 segment buffers.*
$\includegraphics[width=3.2in, height=1.8in]{ins.eps}$ $\includegraphics[width=3.2in, height=1.8in]{res.eps}$ $\includegraphics[width=3.2in, height=1.8in]{harp.eps}$ $\includegraphics[width=3.2in, height=1.8in]{sitar.eps}$ $\includegraphics[width=3.2in, height=1.8in]{uni.eps}$ $\includegraphics[width=3.2in, height=1.8in]{hc.eps}$ $\includegraphics[width=3.2in, height=1.8in]{sf.eps}$ $\includegraphics[width=3.2in, height=1.8in]{tpc.eps}$

It is clear from the figure that the WOLF significantly reduces the overall write cost compared to the LFS. The new design reduces the overall write cost by up to 53%. The overall write cost is most reduced when the disk space utilization is high. When the disk becomes more full, the garbage collection is more important. WOLF plays the more important role in reducing garbage on the disk.

Although the eight traces have different characteristics, we can see that the performance of WOLF is not sensitive to the variation in workloads. This derives from our heuristic reorganizing algorithm. On the other hand, LFS performs especially poor for the TPC-D workload because of its random updating behavior. This is not a surprise. Similar behavior was observed by Seltzer and Smith in [17]. WOLF, on the other hand, significantly reduces the garbage collection overhead so it still performs well under TPC-D.