J. Mogul. Speedier Squid: A case study of an Internet server performance problem. ;login:, 24(1):50-58, February 1999.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....similar to ours at a major Web site. Several studies suggest that disk I O overhead can be a significant percentage of the overall overhead on Web servers and proxy servers [4, 9, 5, tt] Rousskov and Soloviev [tt] observed that disk delay contributes 30 towards total hit response time. Mogul [9] suggests that disk I O overhead of proxy disk caching turns out to be much higher than the latency improvement from cache hits. Kant and Mohapatra [4] point out that, with network bandwidth increasing much faster than server capacity, more and more bottlenecks will be observed on the server side, ....

J. Mogul. Speedier Squid: A case study of an Internet server performance problem. ;login:, 24(1):50-58, February 1999.

....disk allocator so that after restart, the disk allocator can continue where it left off. Fast start is also desirable after a system failure. 1. 1 Related Work Several studies suggest that disk I O overhead can be a significant percentage of the overall overhead on Web servers and proxy servers [4, 9, 5, 10]. Rousskov and Soloviev [10] observed that disk delay contributes 30 towards total hit response time. Mogul [9] suggests that disk I O overhead of proxy disk caching turns out to be much higher than the latency improvement from cache hits. Kant and Mohapatra [4] point out that, with network ....

....a system failure. 1.1 Related Work Several studies suggest that disk I O overhead can be a significant percentage of the overall overhead on Web servers and proxy servers [4, 9, 5, 10] Rousskov and Soloviev [10] observed that disk delay contributes 30 towards total hit response time. Mogul [9] suggests that disk I O overhead of proxy disk caching turns out to be much higher than the latency improvement from cache hits. Kant and Mohapatra [4] point out that, with network bandwidth increasing much faster than server capacity, more and more bottlenecks will be observed on the server side, ....

J. Mogul. Speedier Squid: A case study of an Internet server performance problem. ;login:, 24(1):50--58, February 1999.

....their associated file systems. While UNIX like file systems are widely available and highly reliable, they often result in poor performance for web proxies. For instance, Rousskov and Soloviev [35] observed that disk delays in web proxies contribute about 30 toward total hit response time. Mogul [28] observed that the disk I O overhead of caching turns out to be much higher than the latency improvement from cache hits at the web proxy at Digital Palo Alto firewall. Thus, to save the disk I O overhead the proxy is typically run in non caching mode. These observations should not be surprising, ....

....a significant percentage of their time performing disk I O. Rousskov and Soloviev observed that disk delays contribute as much as 30 towards total hit response time [35] Mogul suggests that disk I O overhead of disk caching turns out to be much higher than the latency improvement from cache hits [28]. Thus, to save the disk I O overhead the proxy is typically run in its non caching mode [28] To reduce disk I O overhead, Soloviev and Yahin suggest that proxies should have several disks [39] in order to distribute the load among them, and that each disk should have several partitions in order ....

[Article contains additional citation context not shown here]

Jeffrey C. Mogul. Speedier Squid: A Case Study of an Internet Server Performance Problem. ;login: The USENIX Association Magazine, 24(1):50--58, 1999.

....and Soloviev [33] observed that disk delays contribute about 30 toward total hit response time. Mogul states that their observations of the web proxy at Digital Palo Alto firewall suggest the disk I O overhead of caching turns out to be much higher than the latency improvement from cache hits [27]. Thus, to save the disk I O overhead the server is typically run in its non caching mode. All the authors are also with the Unversity of Crete. In this paper we study the overheads associated with disk I O for web proxies, and propose secondary storage management alternatives that improve ....

....proxies spend a significant percentage of their time doing disk I O. Rousskov and Soloviev observed that disk delays contribute 30 towards total hit response time [33] Mogul suggests that disk I O overhead of disk caching turns out to be much higher than the latency improvement from cache hits [27]. To reduce disk I O overhead Soloviev and Yahin suggest that proxies should have several disks, and that each disk should have several partitions. Using sophisticated write distribution policies Soloviev and Yahin are able to spread requests over several disks and to cluster requests on the same ....

[Article contains additional citation context not shown here]

Jeffrey C. Mogul. Speedier Squid: A Case Study of an Internet Server Performance Problem. ;login: The USENIX Association Magazine, 24(1):50--58, 1999.

.... considered a fast proxy, but its primary inefficiency lies in the way it misuses the disks [28] which does not significantly affect CPU time (and in any case, our installation does not use a disk cache) Also, we used a modified version of Squid that avoid one particular CPU scaling error [31] and a modified kernel that avoid several CPU scaling problems [3] Although the digest computation can be done off the cache s critical path, a cache participating in duplicate suppression must still maintain and use an additional lookup table. Lookups using a digest instead of a URL would not ....

Jeffrey C. Mogul. Speedier Squid: A Case Study of an Internet Server Performance. ;login: 24(1):50-58, February, 1999.

....not only spends cycles within comm select( but also tramples on most or all of the contents of the data cache, resulting in excessive cache misses for other procedures. A relatively minor modification to the algorithm in comm select( allowed us to eliminate almost all of the cache misses [Mog99] We also found that the procedure, when scanning the bitmaps returned by select( was using an inefficient technique; we replaced it by a faster, albeit somewhat less portable, mechanism. As a result, we eliminated essentially all of the CPU cycles spent in comm select( see Table 13. Not ....

Jeffrey C. Mogul. Speedier Squid: A case study of an Internet server performance problem. In preparation, 1999.

....(usually 2048 bytes, for TCP sockets) The application then scans the result bitmaps to discover the readable or writable file descriptors, and normally invokes handlers for those descriptors. Figure 2 is an oversimplified example of how an application typically uses select( One of us has shown[15] that the programming style used here is quite inefficient for large numbers of file descriptors, independent of the problems with select( For example, the construction of the input bitmaps (lines 8 through 12 of Figure 2) should not be done explicitly before each call to select( instead, the ....

J. C. Mogul. Speedier Squid: A case study of an Internet server performance problem. ;login:, pages 50--58, February 1999.

....(usually 2048 bytes, for TCP sockets) The application then scans the result bitmaps to discover the readable or writable file descriptors, and normally invokes handlers for those descriptors. Figure 2 is an oversimplified example of how an application typically uses select( One of us has shown[15] that the programming style used here is quite inefficient for large numbers of file descriptors, independent of the problems with select( For example, the construction of the input bitmaps (lines 8 through 12 of Figure 2) should not be done explicitly before each call to select( instead, the ....

J. C. Mogul. Speedier Squid: A case study of an Internet server performance problem. Submitted to ;login:, 1999.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC