Some network-picked hpc materials and network-picked hpc Materials

Some network-picked hpc materials and network-picked hpc Materials

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 256 K, private

L3 8 Mb/12 Mb, shared

Cpu-Memory bandwidth: 32 Gb/s

 

Intel Xeon E5-2670

8-core, 2.6 GHz

Cache

L1 Data 32 K, private

L1 Instruction 32 K, private

L2 256 K, private

L3 20 Mb, shared

CPU-Memory bandwidth 51.2 G/s

 

AMD processors

2.2 GHz

Cache

L1 Data 64 k (2-way)

L1 Instruction 64 k (2-way)

L2 512 K private

L3 2 M shared

 

Direct-connect Architecture

CPU-memory bandwidth 10.7 G/s per socket F

Other connect socket bandwidth 8 G/s (2-way)

 

4x Infiniband Interconnect

* SDR 1.25 G/s

* DDR 2.5 G/s

* QDR 5 G/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.