next up previous
Next: Filemark comparison Up: Performance comparison Previous: Performance comparison

Tiobench comparison

Figure 1: tiobench sequential write throughput results comparison
6#6

Figure 2: tiobench sequential read throughput results comparison
7#7

Although there have been a huge number of changes between the Linux 2.4.29 kernel to the Linux 2.6.10 kernel could affect overall performance (both in and outside of the ext3 filesystem), we expect that two ext3 features, removing BKL from ext3 (as described in Section 2.2) and reservation based block allocation (as described in Section 2.3) are likely to significantly impact the throughput of the tiobench benchmark. In this sequential write test, multiple threads are sequentially writing/allocating blocks in the same directory. Allowing allocations concurrently in this case most likely will reduces the CPU usage and improves the throughput. Also, with reservation block allocation, files created by multiple threads in this test could be more contiguous on disk, and likely reduce the latency while writing and sequential reading after that.

Figure 1 and Figure 2 show the sequential write and sequential read test results of the tiobench benchmark, on the three selected kernels, with threads ranging from 1 to 64. The total files size used in this test is 4GB and the blocksize is 16348 byte. The test was done on a single 18G SCSI disk. The graphs indicate significant throughput improvement from the 2.4.29 kernel to the Linux 2.6.10 kernel on this particular workload. Figure 2 shows the sequential read throughput has been significantly improved from Linux 2.4.29 to Linux 2.6.10 on ext3 as well.

When we applied the extents patch set, we saw an additional 7-10% throughput improvement on tiobench sequential write test. We suspect the improvements comes from the combination of delayed allocation and multiple block allocation patches. As we noted earlier, having both features could help lay out files more contiguously on disk, as well as reduce the times to update the metadata, which is quite expensive and happens quite frequently with the current ext3 single block allocation mode. Future testing are needed to find out which feature among the three patches (extents, delayed allocation and extent allocation) is the key contributor of this improvement.


next up previous
Next: Filemark comparison Up: Performance comparison Previous: Performance comparison
Mingming Cao 2005-07-26