[url]http://usenix.org/events/fast04/tech/full_papers/radkov/radkov_html/node14.html[/url]
Impact of Sequential and Random I/O
Two key factors impact the network message overheads of data operations--the size of read and write requests and the access characteristics of the requests (sequential or random). The previous section studied the impact of request sizes on the network message overhead. In this section, we study the effect of sequential and random access patterns on network message overheads.
To measure the impact of reads, we create a 128MB file. We then empty the cache and read the file sequentially in 4KB chunks. For random reads, we create a random permutation of the 32K blocks in the file and read the blocks in that order. We perform this experiment first for NFS v3 and then for iSCSI. Table 4 depicts the completion times, network message overheads and bytes transferred in the two systems. As can be seen, for sequential reads, both NFS and iSCSI yield comparable performance. For random reads, NFS is slightly worse (by about 15%). The network message overheads and the bytes transfered are also comparable for iSCSI and NFS.
Next, we repeat the above experiment for writes. We create an empty file and write 4KB data chunks sequentially to a file until the file size grows to 128MB. For random writes, we generate a random permutation of the 32K blocks in the file and write these blocks to newly created file in that order. Table 4 depicts our results. Unlike reads, where NFS and iSCSI are comparable, we find that iSCSI is significantly faster than NFS for both sequential and random writes. The lower completion time of iSCSI is due to the asynchronous writes in the ext3 file system. Since NFS version 3 also supports asynchronous writes, we expected the NFS performance to be similar to iSCSI. However, it appears that the Linux NFS v3 implementation can not take full advantage of asynchronous writes, since it specifies a limit on the number of pending writes in the cache. Once this limit is exceeded, the write-back caches degenerates to a write-through cache and application writes see a pseudo-synchronous behavior. Consequently, the NFS write performance is significantly worse than iSCSI. Note also, while the byte overhead is comparable in the two systems, the number of messages in iSCSI is significantly smaller than NFS. This is because iSCSI appears to issue very large write requests to the server (mean request size is 128KB as opposed to 4.7KB in NFS).
[img]http://usenix.org/events/fast04/tech/full_papers/radkov/radkov_html/img12.png[/img]
Impact of Sequential and Random I/O
Two key factors impact the network message overheads of data operations--the size of read and write requests and the access characteristics of the requests (sequential or random). The previous section studied the impact of request sizes on the network message overhead. In this section, we study the effect of sequential and random access patterns on network message overheads.
To measure the impact of reads, we create a 128MB file. We then empty the cache and read the file sequentially in 4KB chunks. For random reads, we create a random permutation of the 32K blocks in the file and read the blocks in that order. We perform this experiment first for NFS v3 and then for iSCSI. Table 4 depicts the completion times, network message overheads and bytes transferred in the two systems. As can be seen, for sequential reads, both NFS and iSCSI yield comparable performance. For random reads, NFS is slightly worse (by about 15%). The network message overheads and the bytes transfered are also comparable for iSCSI and NFS.
Next, we repeat the above experiment for writes. We create an empty file and write 4KB data chunks sequentially to a file until the file size grows to 128MB. For random writes, we generate a random permutation of the 32K blocks in the file and write these blocks to newly created file in that order. Table 4 depicts our results. Unlike reads, where NFS and iSCSI are comparable, we find that iSCSI is significantly faster than NFS for both sequential and random writes. The lower completion time of iSCSI is due to the asynchronous writes in the ext3 file system. Since NFS version 3 also supports asynchronous writes, we expected the NFS performance to be similar to iSCSI. However, it appears that the Linux NFS v3 implementation can not take full advantage of asynchronous writes, since it specifies a limit on the number of pending writes in the cache. Once this limit is exceeded, the write-back caches degenerates to a write-through cache and application writes see a pseudo-synchronous behavior. Consequently, the NFS write performance is significantly worse than iSCSI. Note also, while the byte overhead is comparable in the two systems, the number of messages in iSCSI is significantly smaller than NFS. This is because iSCSI appears to issue very large write requests to the server (mean request size is 128KB as opposed to 4.7KB in NFS).
[img]http://usenix.org/events/fast04/tech/full_papers/radkov/radkov_html/img12.png[/img]
740
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
