Multiple Transferred Files

Even though it is likely that only a few files will be in heavy demand by the clients, we should also investigate how the performance of the system is affected when the number of files frequently requested increases. We consider 1 up to 16 different files of 512 MB each, all stored on a single disk of the server, and downloaded with equal probability. We expect skewness in the popularity of different files will only improve the locality for some of them and will be less stressful for the server resources overall. The clients receive data over 10Mbit/s links, and the system load is equal to 40%. In Figure 12(a), we illustrate the network throughput of the server with sequential and out-of-order transfers respectively. In the out-of-order case, the measured throughput remains roughly above 50 MByte/s when we have up to 8 files, and slightly drops to 48MB/s with 16 files. From Figure 12(b), we see the average disk throughput to increase linearly with the number of files when they are up to eight, and reaching 10MB/s at 16 files. This behavior is reasonable because, with more active files, the number of separate disk access streams increases, and the disk throughput approaches 10MByte/s where it starts limiting the network throughput. With sequential transfers, the disk throughput always limits the network throughput and performance only worsens as the number of different files requested increases.

  

Figure: We examine the sensitivity of the system to the block size parameter, when mixing equiprobable requests from clients with 10T and T3 links requesting a file of 512MB at system load 20% and 40%. Both the network throughput and the miss ratio are adversely affected by block sizes smaller than 64KB, but remain insensitive to larger values.

[]

[]

Figure: We investigate the effect of the queue length limit, when mixing equiprobable requests from clients with 10T and T3 links for a 512MB file at different loads. As the queue length limit increases, the disk throughput also grows leading to higher miss ratio. Eventually, the disk bandwidth becomes bottleneck which makes the server network throughput to drop.

[]

[]