Erik Reidel, Catharine van Ingen, and Jim Gray. A Performance Study of Sequential I/O on Windows NT(TM) 4. In Proceedings of the 2 nd USENIX Windows NT Symposium, USENIX, August 3-4, 1998.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....This latter result occurs because NTFS logs metadata changes in memory and delays writing them to disk, while Berkley FFS forces these changes immediately out to the disk. Riedel, van Ingen, and Gray investigated means of optimizing NTFS for access to large sequential files using SCSI disks [14, 15]. The study found that a combination of techniques permitted achievement of half the maximum peak advertised performance of the hardware (which does not take into account system overhead, bus contention, and other issues that must be dealt with in realistic use but that are often ignored by system ....

Erik Reidel, Catharine van Ingen, and Jim Gray. A Performance Study of Sequential I/O on Windows NT(TM) 4. In Proceedings of the 2 nd USENIX Windows NT Symposium, USENIX, August 3-4, 1998.

....of data read or written. The peak file system write throughput was obtained from requests of size 8 KB or up. Also, on average, suppressing file caching gave much more consistent and higher throughput for write operations. Our results are in line with the analysis of NTFS performance presented in [19]. As for reads, throughput is best with requests of size 32 KB or 64 KB and file caching offers slightly higher throughput because of prefetching. However, for large files, turning off the caching provides higher throughput. In our experiments, we turn off the caching and use requests of size 64 ....

....processes belong to the same SMP node or not. Sharing of an SMP between clients and servers hurts performance when multiple processors in a node try to transfer a message. Even if the theoretical peak performance of PCI is 133 MB s, actual available performance is much lower than that [19]. All processors share a PCI bus and network interface, and thus message passing can be bottlenecked by contention for those shared resources. In the system that we used for our experiments, contention reduces the message passing throughput between two processors in the same SMP node to ....

E. Riedel, C. van Ingen, and J. Gray. A Performance Study of Sequential I/O on Windows NT 4. In Proceedings of the Second USENIX Windows NT Symposium, pages 1--10, Seattle, WA, August 1998.

....throughput was obtained from requests of size 8 128 KB, whereas without file caching, bigger request sizes (32 1024 KB) performed better on average and also gave much more consistent performance for write operations. The results are consistent with the analysis of NTFS performance presented in [13]. As for read, throughput is best with file caching turned on, because of the performance advantage offered by prefetching into the file cache. For reads, request sizes of 8 64 KB lead to peak performance. So in our experiments with Panda on the PC cluster, we use 64 KB application read and write ....

....message size and the total amount of data transferred. Sharing of an SMP between clients and servers hurts performance when multiple processors in a node try to send receive a message. Even if the theoretical peak performance of PCI is 133 MB s, real achieved performance is much lower than that [13]. All SMP processes share a PCI bus and thus message passing can be bottlenecked by the PCI bus connecting to a fast network like Myrinet. In the system that we used for our experiments, contention reduces the message passing throughput between two processors in the same node to approximately 40 ....

E. Riedel, C. van Ingen, and J. Gray. A Performance Study of Sequential I/O on Windows NT 4. In Proceedings of the Second USENIX Windows NT Symposium, pages 1--10, Seattle, WA, August 1998.

....to a richer one, would move the theoretical bottleneck to the PCI bus of the cluster machines. 4.1.2 NTFS Input Output Performance As described in Chapter 3, this application is input output intensive; therefore it is very important to tune the reading and writing stages. Riedel et al.[25] studied NTFS performance for sequential input output. In addition to researching the maximum throughput of every input output scheme available through NTFS, they studied the advantages and disadvantages of the Write Cache Enable (WCE) mode, and how to exploit parallelism by software and hardware ....

Erik Riedel, Catharine van Ingen, and Jim Gray. A performance study of sequential i/o on windows nt 4. In Proceedings of 2nd USENIX Windows NT, pages 1--10, 1998.

....but we were not able to control parameters like request size or the placing of data read. The write stage, using buffered I O and small requests was below the expected performance. Therefore we sought for better mechanisms. 5.3 Tuning I O Disk to disk benchmarks are I O intensive. Riedel et al. [16] studied the NTFS performance for sequential I O, they tested different I O schemes, with the exception of memory mapped files. Some of their conclusions advise to make use of using large requests and unbuffered asynchronous I O (which for writing avoids the risks that WCE represents) with several ....

Erik Riedel, Catharine van Ingen, and Jim Gray. A performance study of sequential i/o on windows nt 4. In Proceedings of 2nd USENIX Windows NT, pages 1--10, 1998.

....than one in fifteen of the leaflevel pages in the AM on each query. Blobworld image processing is computeintensive and time consuming, 3 so it is done 1 Which the reader is invited to examine at http: dlp.cs.berkeley.edu photos blobworld . 2 Using Seagate Barracuda ultra wide SCSI 2 drives, [17] measures a throughput of 9 MB s under Windows NT. The average seek time and rotational delay for this drive are 7.1 ms and 4.17 ms, respectively. For 8 KB transfers, this results in a ratio of 14 sequential I Os for each random I O. In the past, raw drive throughput has increased faster than seek ....

E. Riedel. A performance study of sequential i/o on windows nt 4. In Proc. of the 2nd USENIX Windows NT Symposium, Seattle, WA, 1998.

....7 in excess coverage. Essentially, the only problem with a bulk loaded R tree is its sloppy BPs. We also bulk loaded two other previously proposed 4 Which the reader is invited to examine at http: elib.cs.berkeley.edu photos blobworld . 5 Using Seagate Barracuda ultra wide SCSI 2 drives, [20] measures a throughput of 9 MB s under Windows NT. The average seek time and rotational delay for this drive are 7.1 ms and 4.17 ms, respectively. For 8 KB transfers, this results in a ratio of 14 sequential I Os for each random I O. In the past, raw drive throughput has increased faster than seek ....

E. Riedel. A performance study of sequential i/o on windows nt 4. In Proc. of the 2nd USENIX Windows NT Symposium, Seattle, WA, 1998.

....depend to a great deal on the bus characteristics, on whether file system buffering is on, on whether write caching is present, and on the request size, on the file system used, and on whether the operation is a read or a write. These values are roughly indicative of the actual values. See [GG97, RvIG98, GK97] for recent detailed measurements on different disks and file systems. 0 500 1000 1500 2000 1000 2000 3000 4000 5000 6000 number of records in range to be read time (msecs) clustered B Tree LPD Figure 4: Time (milliseconds) take for range searches for LPDs and clustered B Trees 0 5 ....

E. Riedel, C. van Ingen, and J. Gray. A performance study of sequential I/O on Windows NT 4. In USENIX Windows NT Symposium, 1998.

....or SR tree perform in total. R tree and SR tree performance is comparable, with the spheres in the SR tree BPs saving a small amount of leaf level excess coverage loss relative to the R tree. 6 However, for both R trees and SR trees, about 30 4 Using Seagate Barracuda ultra wide SCSI 2 drives, [19] measures a throughput of 9MB s under Windows NT. The average seek time and rotational delay for this drive are 7.1 ms and 4.17 ms, respectively. For 8KB transfers, this results in a ratio of 14 sequential I Os for each random I O. In the past, raw drive throughput has increased faster than seek ....

E. Riedel. A performance study of sequential i/o on windows nt 4. In Proc. of the 2nd USENIX Windows NT Symposium, Seattle, WA, 1998.

....the storage size. Note, though, that the threshold result set size does not double as well. In contrast to uniformly 10 This test assumes that total execution time of the workload under consideration is dominated by page access cost. 11 Using Seagate Barracuda ultra wide SCSI 2 drives, Rie98] measures a throughput of ca. 9MB s under Windows NT. The average seek time and rotational delay for this drive are 7.1ms and 4.17ms, respectively. For 8KB transfers, this results in a ratio of 14 sequential I Os for each random I O. In the past years, raw drive throughput has increased faster ....

Erik Riedel. A Performance Study of Sequential I/O on Windows NT 4. In Proc. 2nd USENIX Windows NT Symposium, Seattle, WA, 1998.

No context found.

Erik Riedel. A Performance Study of Sequential I/O on Windows NT 4. In Proc. 2nd USENIX Windows NT Symposium, Seattle, WA, 1998.

No context found.

Erik Riedel. A Performance Study of Sequential I/O on Windows NT 4. In Proc. 2nd USENIX Windows NT Symposium, Seattle, WA, 1998.

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