2 ms·
Performance Analysis Guide for Xeon 5500 (2009, pdf)
- dragontamer 8y agoWhile the Xeon 5500 / Nehelem processor is 9-years-old at this point, this document meticulously goes through its architecture... as well as a huge number of tips on how to use hardware performance counters to guide optimization of code. If anyone knows of a similarly good document with modern architectures (ie: Intel's Skylake / 6th Gen i7 or later or AMD Zen), I'm all ears. But this document looks to be exceptionally good, despite its age. ----------- For the uninitiated: modern systems have performance counters available through tools like Linux's perf, which can count branch prediction misses, or cache-misses... information which can guide a low-level programmer towards optimizations. Figuring out "why" your program is slow and how to change your algorithms to improve speed requires an understanding of computer architecture (ex: rearranging a for-loop to be L1 cache friendly can lead to 10x speed increases, as you go from main-memory constrained to L1 cache constrained). But in the real world, when dealing with real problems, it is difficult to figure out performance issues. And that's where knowledge of hardware counters can come in.
- wyldfire 8y agoI did a lot of work with the 5600 Westmere. After working out the "easy" stuff related to NUMA and IOH, the (relatively new IIRC) uncore was a big remaining part of the recipe for performance when trying to saturate the PCIe devices.
- dragontamer 8y agoInteresting story! Care to elaborate on what kind of PCIe devices you were using? Was it GPGPU compute? Or were you trying to saturate a router / network of some kind? And do you remember what kind of problems you faced? (Latency or bandwidth??)
- wyldfire 8y agoThey were Intel X520-SR2 Niantic 10GbE cards (later 82599ES). Bandwidth. It was challenging to find a stable recipe on linux to receive TCP streams at or near the media limit (x four ports, simultaneously).