"... In PAGE 7: ... In particular we consider the aggregation of MPEG traces in three different scenarios, varying the number of traces and the traces themselves. Table1 sum- marizes the real traces used as well as their average and peak-rate. Using the estimated parameters in the table, we compute the optimal token bucket descriptors for each aggregated traffic, and compare them to the analytic model; the com- parison criterion being the effectiveness of analytic and empirical parameters in matching the actual LBAP con- straints.... ..."

"... In PAGE 58: ... The TPC-C trace is a transaction processing trace with 14 warehouses, the Web Search trace a disk access trace collected from a web search engine, and the Financial trace collected from enterprise financial applications. Table1 lists the characteristics of the three traces. Some synthetic traces that simulate random disk accesses are also used in the comparison of the CVC and CFVC schedulers.... In PAGE 63: ... In each test setup, three virtual disks, VD0, VD1, and VD2, are hosted by a physical disk. VD0 and VD1 are configured as real-time virtual disks running real traces shown in Table1 and each reserve 50% of the disk bandwidth. VD2 runs as a best-effort virtual disk with a random trace.... In PAGE 110: ... Small sized requests ( lt; 16 KB) dominate. Read Path Write Path Hit Miss Overwritten Flushed NFS server 2 3 1 2 kHTTPd 1 2 N/A N/A Table1 0: Number of data copying operations per request in NFS server or kHTTPd for reading data or writing data. kHTTPd copies data directly from file system buffer cache to network stack when serving requests and thus for each request incurs one less copying operation than NFS server in the read path.... In PAGE 117: ... In general, the performance gain of kHTTPd-NCache over kHTTPd-original is smaller than that of NFS-NCache over NFS-original as shown in Figure 32(b) for two reasons. First, the data copying overhead per request in kHTTPd is inherently lower than that in NFS server, as shown in the read path column of Table1 0, because of the sendfile interface. Second, the per-packet overhead of HTTP is higher than that of NFS because HTTP runs on TCP and NFS runs on UDP in our experiments.... In PAGE 118: ...NETWORK-CENTRIC BUFFER CACHE ORGANIZATION 98 Request Size Without NCache With NCache 4 KB 3 packets 4-5 packets 8 KB 6 packets 7-8 packets 16 KB 12 packets 13-14 packets 24 KB 17 packets 19-20 packets 32 KB 23 packets 25-26 packets Table1 1: Average number of packet transmissions for each request size in NFS. In NCache case, more packets need to be transmitted because of mis-alignment.... ..."

"... In PAGE 88: ... The TPC-C trace is a transaction processing trace extracted from 14 warehouses, the Web Search trace a disk access trace collected from a web search engine, and the Financial trace collected from enter- prise financial applications. Table2 lists the characteristics of the three traces. The TPC-C traces and Financial traces have latency requirements but the Web Search trace does not.... In PAGE 89: ...AVAILABILITY SUPPORT FOR VIRTUAL DISKS 69 (E) for all VDs is set to 98%, and the input workloads are the Financial traces and the Web Search traces. The Financial workload has a delay requirement as shown in Table2 and is latency-bound while the Web Search workload has no delay require- ment and is throughput-bound. All traces have been scaled down with a factor of 4 to allow more VDs to be admitted.... In PAGE 90: ... Two types of semi-synthetic workloads are used, which are derived from the TPC-C traces and the Web Search traces. The semi-synthetic traces based on the TPC-C workloads have a delay requirement as shown in Table2 and are latency-bound while the traces based on the Web Search workloads have no delay requirement and are throughput-bound. All traces have been scaled down with a factor of 4 to allow more VDs to be admitted.... ..."

"... In PAGE 9: ... All simulations assume two-speed disks and 8 (PDC), 9 (MAID-1+), or 10 (MAID-2+) disks in the array. Table7 lists the energy and percentage of requests delayed (by more than 200 milliseconds) for each technique and trace after their warm up periods. EO consumes 2022095 J (HUM) and 256745 J (POP) with virtually no delayed requests.... ..."

"... In PAGE 9: ... All simulations assume two-speed disks and 8 (PDC), 9 (MAID-1+), or 10 (MAID-2+) disks in the array. Table7 lists the energy and percentage of requests delayed (by more than 200 milliseconds) for each technique and trace after their warm up periods. EO consumes 2022095 J (HUM) and 256745 J (POP) with virtually no delayed requests.... ..."

"... In PAGE 10: ... We tried several param- eter settings for both PD-50 and PD-1, but failed to ob- tain stability. Table1 compares the HS controller with the PD controllers (since in this case both PD controllers per- form similarly, we only show one set of values). The re- ward numbers represent accumulated reward values (total throughput minus scaled total delay).... ..."