PC-3200 vs PC-5300 benchmarking issue

[moorek]

I have a customer that is benchmarking two identical servers, where one has been upgraded to newer processors and memory. He is finding that the latency on the newer components is worse based on the benchmark test.

The older server has E7320 Xeon processors and DDR2-400 (PC-3200) memory. The new server has E7420 Xeon processors and DDR2-667 (PC-5300) memory.

The benchmark he is running is called Latency from cpuid.com. My guess is that the results are skewed since the PC-5300 memory has L3 cache and the PC-3200 memory does not so it is an apples to oranges comparison. Can anyone shed some light on this? Thanks!


E7320 with DDR2-400 (PC-3200) memory
stride 4 8 16 32 64 128 256 512
size (Kb)
1 4 4 4 4 4 4 4 4
2 4 4 4 4 4 4 4 4
4 4 4 4 4 4 4 4 4
8 4 4 4 4 4 4 4 4
16 4 4 4 4 4 4 4 4
32 4 4 4 4 4 4 4 4
64 4 5 7 11 19 19 19 19
128 4 5 7 11 19 19 19 19
256 3 4 5 9 14 14 14 14
512 3 4 5 9 14 14 14 14
1024 3 4 5 9 14 14 14 14
2048 3 4 5 9 15 35 23 24
4096 3 5 5 23 41 249 249 250
8192 3 4 5 23 48 253 249 250
16384 3 4 5 23 70 254 252 250
32768 3 4 5 23 48 253 253 254
2 cache levels detected
Level 1 size = 32Kb latency = 4 cycles
Level 2 size = 2048Kb latency = 17 cycles


E7420 with DDR2-667 (PC-5300) memory
stride 4 8 16 32 64 128 256 512
size (Kb)
1 3 3 3 3 3 3 3 3
2 3 3 3 3 3 3 3 3
4 3 3 3 3 3 3 3 3
8 3 3 3 3 3 3 3 3
16 3 3 3 3 3 3 3 3
32 3 3 3 3 3 3 3 3
64 3 4 5 9 15 15 15 15
128 3 4 5 9 15 15 15 15
256 3 4 5 9 15 15 15 15
512 3 4 5 10 15 15 15 15
1024 3 4 5 9 15 15 15 15
2048 3 4 5 9 15 15 16 17
4096 3 4 5 10 18 94 94 95
8192 3 4 5 10 21 115 110 111
16384 3 4 5 10 21 380 328 374
32768 3 4 5 10 23 380 380 381
3 cache levels detected
Level 1 size = 32Kb latency = 3 cycles
Level 2 size = 2048Kb latency = 15 cycles
Level 3 size = 8192Kb latency = 103 cycles

 

[Emulex]

some of those fancy motherboards actually add another level of cache for coherency between nodes that could add some overhead. (unisys/ibm 3950M2/etc).

i'd also read the specs - many motherboards clock down their memory at full load (say 64 4gb sticks) due to the electrical load it places on the machines.

bout tired of dunnington lol - dual nehalem can smoke a quad dunnington. the nehalem-ex machines (hp/ibm) should be out by summer 🙂

biggest machines (in one node) i've ever seen.

 

[SuRgEoN]

"DDR2-400 (PC-3200)" This is actually DDR not DDR2.

 

[tcsenter]

He is finding that the latency on the newer components is worse based on the benchmark test.

Well then you mixed up the results or something, because those results show the E7420 with DDR2-667 to have better performance nearly across the board, except for the L3 cache latency, which is obviously going to be much worse than a CPU without L3 (no L3 = nothing to test).

"DDR2-400 (PC-3200)" This is actually DDR not DDR2.

PC2-3200 = DDR2-400.