####Author: Gavin Baker ####Contact: gavin.m.baker@gmail.com gabaker@calpoly.edu

##TARUC Benchmark Overview:

CUDA PCIe Bandwidth, Memory and Multi-GPU Benchmarks that accounts for multi socket systems with many variable model GPUs

##Package Requirements:

• C++11
• CUDA 7.5 (untested on earlier)
• NVML
• HWLOC
• python2.7
• OpenMP (omp.h)
• Unix operating system

##Features:

• CUDA support for multi-gpu systems
• Support for diverse topology recognition using hwloc, numa.h and cpuset.h
• matplot lib plotting of complete benchmark runs using provided parameter file

##Non-Features:

• Distributed system awareness
• Network communicaton

##TARUC Micro-Benchmarks:

###Memory Management Overhead

###Memory Transfer Bandwidth

####Host-Host Transfer Bandwidth (single-threaded)

####Host-Device Transfer Bandwidth (single-threaded)

####Device-Device Transfer Bandwidth (single-threaded)

###Resource Contention

####Local Host Memory Bandwidth (multi-threaded)

####Inter-Socket (QPI) Host Memory Bandwidth (multi-threaded)

####Single Device PCIe Memory Bandwidth (multi-threaded)

####GPU Pair Memory Bandwidth (multi-threaded)

####Single Host Multi-Device Memory Bandwidth (multi-threaded)

###NUMA Latency

##TARUC Run Instructions:

Run w/ Default parameters:

  TARUC_Bench/scripts/run_benchmark.sh parameters.in

Run w/ provided parameter file (parameters.in)

  TARUC_Bench/scripts/run_benchmark.sh parameters.in

##Understanding Benchmark Results

##Parameter Explaination

The following is a sample parameter file demonstrating correct format.

------- Benchmark Parameters --------
Run Tag = anatta
------------- All Tests -------------
Use All GPUs = true
Test All Mem Types = true
Run Burst Tests = false
Run Range Tests = true
Run Sustained Tests = true
Run Socket Effect Tests = true

Burst Block Size (bytes) = 64000000
---------- Memory Overhead ----------
Run Memory Overhead Test = false
Range Min = 100
Range Max = 2500000000
--------- Host-Host Bandwidth -------
Run Host-Host Tests = false
Vary Access Patterns = true
Range Min = 100
Range Max = 2500000000
------- Host-Device Bandwidth -------
Run Host-Device Tests = false
Vary Access Patterns = true
Range Min = 100
Range Max = 2500000000
----------- P2P Bandwidth -----------
Run P2P Tests = false
Range Min = 100
Range Max = 1500000000
-------- Resource Congestion --------
Run Congestion Test = true

##Directory Structure and Project Files:

Folders:

• scripts/
• results/
• misc/
• sample_images/

Files:

• benchmark.cu
• benchmark.h
• helper_cuda.h
• helper_string.h
• parameters.cpp
• parameters.h
• parameters.in
• topology.cu
• topology.h
• timer.cu
• timer.h
• nvml.h
• Makefile