R Lists06
2007-Apr-07 16:34 UTC
[CentOS] OT: general question re processor, l2 and l3 cache etc
Greetings Please forgive the OT question yet I highly value the experience and wisdom on this list I am wondering if anyone here can address the performance difference between having a processor board with say 256KB L2 *and* 2048KB L3 cache *VERSUS* just having the same processor board with just the L2 cache in a centos server environment... Please figure that all other necessary and related componentry is well above needed amounts and high performance specs for a professional server environment. Thanks in advance!! - rh -- Abba Communications - Internet PO Box 7175 Spokane, WA 99207-7175 www.abbacomm.net
John R Pierce
2007-Apr-07 18:38 UTC
[CentOS] OT: general question re processor, l2 and l3 cache etc
R Lists06 wrote:> Greetings > > Please forgive the OT question yet I highly value the experience and wisdom > on this list > > I am wondering if anyone here can address the performance difference between > having a processor board with say 256KB L2 *and* 2048KB L3 cache *VERSUS* > just having the same processor board with just the L2 cache in a centos > server environment... > > Please figure that all other necessary and related componentry is well above > needed amounts and high performance specs for a professional server > environment. >Thats absolutely unanswerable. many servers are far more network and/or disk IO bound than they are CPU bound, so that extra cache would do nothing. even on a pure compute workload, many inner loops fit almost entirely into the L2 cache and the L3 cache wouldn't be doing much at all. even if you are getting lots of cache hits from this L3 cache, the actual benefit would depend on how slow the external memory latency is vs. this L3 cache. today's modern CPUs often have 1-4MB of L2 cache, so L3 is even more pushed out... larger caches are a classic case of diminishing returns The only way you'll get meaningful data on this would be to take these two systems, configured identically, and run the same application workloads on them, and measure the performance.