My new MP7,1 16-core came with its stock 4x 8GB R-DIMMs. I replaced them with 12x 32GB R-DIMMs.
The 16-core has some 22 MB for its L3 Cache, and I modified the STREAM's code to ensure it had its arrays 4x the L3 Cache size.
I ran STREAM for each memory configurations... see attachments.
Summary
4x 8GB gave: Triad 50.205 GBytes/sec - Stock
12x 32GB gave: Triad 84.568 GBytes/sec - 1.7x Stock
If your Sonnet PCIe x16 card is delivering some 8 GB/s then comparing to using RAM (memory) for reading/writing with the 12x 32GB R-DIMMs installed, one benefits by an order of magnitude. That is, having gobs of RAM for the kernel file cache then reading/writing needed file data to this cache will be extremely fast. Lots of RAM is just plain good, just as speed is good. 👍
In addition out of curiosity I modified STREAM to measure the L3's (22MB in size) memory bandwidth and it averaged 126 GBytes/sec. 👍
The 16-core has some 22 MB for its L3 Cache, and I modified the STREAM's code to ensure it had its arrays 4x the L3 Cache size.
I ran STREAM for each memory configurations... see attachments.
Summary
4x 8GB gave: Triad 50.205 GBytes/sec - Stock
12x 32GB gave: Triad 84.568 GBytes/sec - 1.7x Stock
If your Sonnet PCIe x16 card is delivering some 8 GB/s then comparing to using RAM (memory) for reading/writing with the 12x 32GB R-DIMMs installed, one benefits by an order of magnitude. That is, having gobs of RAM for the kernel file cache then reading/writing needed file data to this cache will be extremely fast. Lots of RAM is just plain good, just as speed is good. 👍
In addition out of curiosity I modified STREAM to measure the L3's (22MB in size) memory bandwidth and it averaged 126 GBytes/sec. 👍
Attachments
Last edited:
