Our micro-benchmark experiments revealed two important characteristics of modern local file systems --aggressive meta-data caching, which benefits meta-data reads, and update aggregation, which benefits meta-data writes. Recall that, update aggregation enables multiple writes to the same dirty block to be ``batched'' into a single asynchronous network write. We explore this behavior further by quantifying the benefits of update aggregation and caching in iSCSI.
We choose eight common operations that read and update meta-data, namely creat,
link, rename, chmod, stat, access, write and mkdir. For each operation,
we issue a batch of 
consecutive calls of that operation and measure the network message overhead of the
entire batch. We vary from 1 to 1024 (e.g., issue 1 mkdir, 2 mkdirs, 4 mkdirs and so on,
while starting with a cold cache prior to each batch). Figure 3
plots the amortized network message overhead per operation for varying batch sizes.
As shown, the amortized overhead reduces significantly with increasing batch sizes, which
demonstrates that update aggregation can indeed significantly reduce the number of network
writes. Note that some of the reduction in overhead can be attributed to meta-data caching in
iSCSI. Since the cache is warm after the first operation in a batch, subsequent operations do not
yield additional caching benefits--any further reduction in overhead is solely due to
update aggregation.
In general, our experiment demonstrates applications that exhibit meta-data locality can
benefit significantly from update aggregation.
 |