1. Background
I studied to improve the computing performance with Roofline model, Regression Predication.
The main idea to improve the performance is Data migration. The KNL (= Kight Landing Processor) processor has two types cache memories. the first one is DRAM that we usually used in personal computer and the last one is HBM (= High Bandwidth Memory). Data migration means to migrate the data between DRAM and HBM as the features.
2. Environment
System : Linux based Knight Landing Processor
Language : Python, Bash Shell Script
Framework & Library : numactl, Intel Advisor
3. Technical Detail
1. Cache Coherence
I had not ever been consider about the cache coherence. but after this study, I have realized the mechanism of it. It has two ways how keep the cache coherence. Snoopy Protocol firslty and the second one is Directory based Protocol.
KNL has a many cache node so Snoopy Protocol is not good choice for this one. It will occur bottlenecks.
2. NUMA structure of KNL processor
KNL memory structure is not UMA (= Uniform Memory Access) unless personal computer. this super computer consist of NUMA sturucture that it means cache node and the processor are dependent each.
3. Predication with Regression Model
I use this regression model to decide whether the data needs to migrate or not.
[논문] HyDM: Data Migration Methodology for Hybrid Memories.pdf
[연구자료] Cache Coherence Protocols.pptx