Check out the new USENIX Web site. next up previous
Next: Impact of Sequential and Up: Micro-benchmarking Experiments Previous: Impact of Directory Depth

Impact of Read and Write Operations

Our experiments thus far have focused on meta-data operations. In this section, we study the efficiency of data operations in NFS and iSCSI. We consider the read and write system calls and measure their network message overheads in the presence of a cold and a warm cache.

To measure the read overhead, we issue reads of varying sizes--128 bytes to 64 KB--and measure the resulting network message overheads in the two systems. For the warm cache, we first read the entire file into the cache and then issue sequential reads of increasing sizes. The write overhead is measured similarly for varying write sizes. The cold cache is emulated by emptying the client and server caches prior to the operation. Writes are however not measured in warm cache mode--we use macro-benchmarks to quantify warm cache effects.

Figure 5 plots our results. We make the following observations from our results. For read operations, iSCSI requires one or two extra messages over NFS to read or update uncached file meta-data (e.g., inode blocks). While NFS incurs a smaller overhead for small cold reads, the read overhead exceeds that of iSCSI beyond 8KB requests. For NFS v2, this is due to the maximum data transfer limit of 8KB imposed by the protocol specification. Multiple data transfers are needed when the read request size exceeds this limit. Although NFS v3 eliminates this restriction, it appears that the Linux NFS v3 implementation does not take advantage of this flexibility and uses the same transfer limit as NFS v2. Consequently, the cold read overhead of NFS v3 also increases beyond that of iSCSI for large reads. In contrast, the NFS v4 implementation uses larger data transfers and incurs fewer messages. In case of the warm cache, since the file contents are already cached at the client, the incurred overhead in NFS is solely due to the consistency checks performed by the client. The observed overhead for iSCSI is due to the need to update the access time in the inode.

Similar observations are true for write requests (see Figure 5(c)). Initially, the overhead of iSCSI is higher primarily due to the need to access uncached meta-data blocks. For NFS, all meta-data lookups take place at the server and the network messages are dominated by data transfers. The network message overhead for NFS v2 increases once the write request size exceeds the maximum data transfer limit; the overhead remains unchanged for NFS v3 and 4.

Figure 5: Network message overheads of read and write operations of varying sizes.
... reads & (b) Warm reads & (c) Cold Writes

next up previous
Next: Impact of Sequential and Up: Micro-benchmarking Experiments Previous: Impact of Directory Depth