"... In PAGE 3: ... Instead, the interrupt is delayed for up to a given amount of time (the interrupt moderation period) in hopes of other packets arriving during that time can be served by the same interrupt. Table2 shows how interrupt coalescing affects CPU utilization, thus increasing the network throughput. TCP/IP packets are 1500 bytes for all measurements.... ..."

"... In PAGE 2: ... A virtualized system cannot deliver performance beyond what the physical hardware is capable of providing after subtracting the overhead required to manage the virtualization. Table1 contains a list of major components that may be either virtual or physical. The capacity or capability of each physical component may be specified by the hardware manufacturer or measured by some accepted methodology.... In PAGE 2: ... Studies have shown that the measurements taken at the logical component level may differ significantly from corresponding measurements at the Virtualization Manager level [BD]. In this paper, we assume that all measurements for the performance metrics defined in Table1 are collected at the Virtualization Manager level. Suppose a virtualized system has n partitions or guests, n G G G , , , 2 1 L , running on a physical system whose four hardware components are Processor P, Memory M, I/O subsystem D, and Network N.... ..."

"... In PAGE 11: ... FPGAs reduce costs in other ways, include shrinking total area, design time, and system complexity. Finally, Table5 shows the utilization of logic blocks and i/o blocks for each device. We present these values separately as each may serve as a limiting factor in FPGA use.... In PAGE 14: ... When examining candidate devices, pin counts may become more important. As shown in Table5 , 6 of the 9 devices were i/o-limited rather than logic limited. We also found it difficult to combine logic efficiency with pin efficiency; in no case were we able to obtain 90% utilization or higher of both logic modules and i/o.... ..."

"... In PAGE 6: ... Table 2 and table 3 shows buffer utilizationfor the Xcent routers and nodes. It appears from Table3 that the Xcent buffer utilization is not high to be a performance limit in various disk cache miss ratios. On the other hand,data buffer on the IO node(ION) can be a bottleneck when we have enough number of disks per node.... ..."

"... In PAGE 13: ...ith an average of only 0.24 Mb of I/O per GB of storage. Based on these results, we project the amount of processing power and storage space that will be needed to drive various types of I/O interconnect to 50% utilization. The results are summarized in Table9 . Note that all the modern I/O inter- connects offer Gb/s bandwidth.... ..."

"... In PAGE 8: ...Table1 list the FPGA resources used in each of the four con gurations. The resources are divided into four categories: ip- ops, FPGA programmable cells, user programmable Input/Output pins, and interconnection busses.... ..."

"... In PAGE 11: ... Based on these results, we project the amount of processing power and stor- age space that will be needed to drive various types of I/O interconnect to 50 percent utilization. The re- sults are summarized in Table6 . Note that all the modern I/O interconnects offer Gb/s bandwidth.... ..."