CALIPER-FRAMEWORK CUDA Version

What is Caliper?

Caliper is an OpenSource framework to compute MTTF (Mean Time To Failure) of a multicore homogeneous grid of cores (MultiCore CPU). This framework was developed by a Research Team in Politecnico di Milano and it's Sequential implementation can be fount at: https://github.com/D4De/caliper

How Caliper Works?

Caliper use Montecarlo Algorithm to approximate MTTF value within a certain "confidence interval".

Scope of our caliper-gpu:

Sequential implementation of caliper has some time limitation since it become unfeasible to simulate big grid of cores. A grid of core 42 * 42 take more than 3 hours to give the resulting MTTF on a Intel I5 8thGen. Our study aim to adapt their algorithm to a CUDA capable gpu so to get results within minutes instead of hours. We explored all possible CUDA implementation and documented major pros/cons of them.

On our Documentation Slides you can find more details about it.

What have we done?

As a project for the GPU & Heterogeneous system course, we created a CUDA version of Caliper. The final version had a 72x speedup [Intel I5 8thGen w.r.t a Tesla K80 on AWS].

Compile Instruction:

Simply compile using "make"

This may require open the makefile and change the CAPABILITY field to the target GPU CUDA CAPABILITY (eg: for a Tesla K80 -> capability is 3.7 -> we put 37)

Framework usage Instructions:

./caliper-gpu [Col][Row][MinCore][Initial Workload][-g <version>]

• COL : Num of Columns in the Cores Grid
• ROW : Num of Rows in the Cores Grid
• MinCore : Num of cores that must be alive to continue simulation
• Initial Workload : Initial Workload of each core (equaly distributed among cores)
• -g : Is for executing gpu version of caliper

GPU VERSIONS: [Best one is 2]

1. Parallel Reduction Basic
2. Coalesced Memory Acces
3. Dynamic Programming
4. Static 2D Grid of thread
5. Struct Coalesced Memory Acces 6-9. ....(other versions)