http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
Improvement of Bonding Driver and Application to Parallel Processing on Commodity PC Cluster
Takafumi Fukunaga 제어로봇시스템학회 2009 제어로봇시스템학회 국제학술대회 논문집 Vol.2009 No.8
We have studied various methods for improving the performance of PC clusters. Most of them have used special hardware, non standard protocols, so that need special added library and complex programming, tending to raise their cost and to spoil their flexibility. To solve their problems we propose improving of Bonding driver. This improvement have a good influence on some kinds of parallel applications on commodity PC clusters without costing a great deal, using comparatively simple method where distribute sending frames to multiple Network Interface Cards (hereafter NICs) on a round-robin basis. This means divide communication loads equally between multiple NICs attached to each PC. We measured the maximum bandwidth and parallel performance for the PC cluster with Gigabit Ethernet. A result shows our proposal method improves the bandwidth and the performance of several parallel benchmarks in comparison with the system using single NIC (hereafter 1NIC) and existing Bonding driver (hereafter Bond). The bandwidth comes up to 1,249Mbps which is 1.33 times and 1.47 times faster than those of 1NICand Bond. The bandwidth of Bond cannot even reach that of 1NIC. The results of SP and BT in NAS Parallel Benchmarks (NPB) come up to 1,715 Mop/s and 5,264Mop/s which are 1.81 and 1.63 times faster than 1NIC, 1.17 and 1.15 times faster than Bond respectively.
Improvement of parallel processing performance by using two kinds of Huge Page
Takafumi Fukunaga,Toshinori Sueyoshi 제어로봇시스템학회 2008 제어로봇시스템학회 국제학술대회 논문집 Vol.2008 No.10
A main memory capacity in PC continues to increase because the memory requirements which applications need get bigger and bigger. It is easy to expand main memory. But TLB is very scarce as before. Under these circumstances, the uses of Huge Page reduce TLB misses by reducing the numbers of needed entries of TLB, consequently give applications which need large memories a good effect in regard to execution speed. But the uses of Huge Page cause large useless memory because operating system round up used memory to a multiple of Huge Page size. Then we propose a method which can use Huge Page while saving memories on IA-64. This proposed method allocates another smaller Huge Page which we added this time to a fragment in the back part of used memory. As a result of experimentation, proposed method show the same speed as the original system and good results in regard to memory requirements. This time we applied this proposed method to PC cluster as parallel processing programs need large memories generally and measured parallel processing performance.