From: Peter W. <wp...@us...> - 2002-10-25 14:43:51
|
I used a very heavy database workload to evaluate two kernels: (1) 2.5.44-mm2: 2.5.44 + Andrew Morton's mm2 patch (2) 2.5.44-mm2-less-shpte: 2.5.44 + Andrew Morton's mm2 patch - Dave McCracken's shpte-ng.patch Andrew' mm2 patch can be found at: http://www.zipworld.com.au/~akpm/linux/patches/2.5/2.5.44/2.5.44-mm2/. Note that Dave's shared pte patch is included in Andrew's mm2 patch. The objective of this comparison is to measure the impact of Dave's patch. ---------------------------------------------------------------------- System Information: - 8-way 700MHz Pentium III Xeon processors, 2MB L2 Cache - 4GB RAM - 6 SCSI controllers connected to 120 9.1 GB disks with 10K RPM ---------------------------------------------------------------------- Based upon the performance metrics and CPU utilization, the two kernels achieve similar performance. However, memory consumption by rmap and page tables is reduced significantly by using Dave's patch. Basically, a large number of I/O prefetchers and database agents can share two small sets of page tables, which in turn reduce the amount of rmap memory. 2.5.44-mm2 2.5.44-mm2-less-shpte ============================================================== pte_chain (max) (MB) 1.2 182.3 PageTables (max) (MB) 5.2 153.0 In total, we save ~330 MB in the low memory area. Regards, Peter Peter Wai Yee Wong IBM LTC Performance Team email: wp...@us... |