source from: https://computing.llnl.gov
Factors determines a large-scale program‘s performance
4 * Application related factors:
5 * algorithms
6 * dataset size
7 * Memory Usage Pattern
8 * Use of IO
9 * Communication Patterns
10 * Task Granularity
11 * Load Balancing
12 * Amdahl‘s Law
13
14 * Hardware factors
15 * Processors Architecture
16 * Memory Hierarchy
17 * I/O configuration
18 * Network
19
20 * Software factors
21 * OS
22 * Compiler
23 * Preprocessor
24 * Communication protocols
25 * Libraries
Performance analysis:
Timers, Profiles, system stat, memory tools
Learn some about hardware archiecture:
Intel Xeon 5500/5600
4-core/ 6-core
2.4/2.8 GHz
Cache
L1 Data 32Kb, private
L1 Instruction 32Kb, private
L2 256K, private
L3 8Mb/12Mb, shared
Cpu-Memory bandwidth: 32 Gb/s
Intel Xeon E5-2670
8-core, 2.6GHz
Cache
L1 Data 32K, private
L1 Instruction 32K, private
L2 256K, private
L3 20Mb, shared
CPU-Memory bandwidth 51.2G/s
AMD processors
2.2 GHz
Cache
L1 Data 64k (2-way)
L1 Instruction 64k(2-way)
L2 512K private
L3 2M shared
Direct - connect Architecture
CPU-memory bandwidth 10.7G/s per socket F
other connect socket bandwidth 8G/s(2-way)
4x Infiniband Interconnect
* SDR 1.25G/s
* DDR 2.5G/s
* QDR 5G/s
Learn something about NUMA
-physical: each node has sevearl(2-4) sockets, each socket has sevearl(4-8) CPU cores. On same socket, cores share L3 cache; socket-socket communcation through CPU-memory bus, almost 2x ~ 5x slower.
-design consideration: CPU affinity(numactl --cpunodebind), local memory policy. other compiler/running-time options(mpirun --bind-to-socket -bynode)
Finally and most importantly, a good algorithm.