Effect of Double Synchronous Writes

Figure 12: Effect on TPC-C throughput as we^M write to two disks synchronously. The total number of disks is a^M constant (36). Labels that include the word ``Delay'' denote^M experiments that propagate all but the first replicas in the^M background. Labels that include the word ``Sync'' denote experiments^M that write two replicas synchronously.

In the previous experiments, all but the first replicas are propagated in the background. To raise the degree of reliability, one may desire to have two copies physically on disk before a write request is allowed to return. We now study the effect of increasing foreground writes on three alternative configurations: a RAID-10, a Doubly Distorted Mirror (DDM), and a $2 \times 9 \times 2$ EW-Array. The DDM performs two synchronous eager-writes and moves one of the two copies to a fixed location in the background. We note that the system that we have called the DDM is in fact a highly optimized implementation that is based on the MimdRAID disk location prediction mechanism and the eager-writing scheduling algorithms, features not detailed in the words of ``write anywhere'' in the simple original simulation-based study [19]. (We do not consider SR-Arrays because the pure form of an SR-Array involves only intra-disk replication which does not increase the reliability of the system.) As expected, extra foreground write slows down both the RAID-10 and the EW-Array. However, for a given request arrival rate that does not cause performance collapse, the response time degradation experienced by the RAID-10 is more pronounced than that seen on the EW-Array. This is because the cost of an extra update-in-place foreground write is relatively greater than that of an extra foreground eager-write. The performance of the DDM lies in between , because the two foreground writes enjoy some performance benefit of eager-writing but the extra update-in-place write becomes costly, especially when the request arrival rate is high. One of the purposes of this third update-in-place write is to restore data locality that might have been destroyed by the eager-writes. This is useful for workloads that exhibit both greater burstiness and locality. Unfortunately, the TPC-C workload is such that it does not benefit from this data reorganization.