cwpearson / cuda-experiments

cuda-experiments

Quick Start

./configure && make && make bench

Configuring

TO auto-detect CC version and create config.mk, try

`./configure`

Otherwise, create config.mk and set the variable CUDA_CC to the version of code you want to generate. For example

CUDA_CC=35

Building

Build all supported tests

make

Running

Run all supported tests and dump outputs in $MODULE.csv

make bench

Run individual tests, with outputs to stdout:

<module>/main

stream-thread

https://devblogs.nvidia.com/maximizing-unified-memory-performance-cuda/

stream-warp

https://devblogs.nvidia.com/maximizing-unified-memory-performance-cuda/

Other Benchmarks

mgbench

blackjackbench

cuda-experiments's Issues

atomics performance

number of conflicting threads in each bin

stride between bins?

cudaMemcpy may return before final DMA is complete

Figure out access counters for triggering migration

use cudaMemAdvise to establish a direct mapping.

Access more and more until migration occurs.

How does cudaMemAdvise preferred location affect this?

page migration throughput

memcpy
prefetch
on-demand (stream?)
on-demand

pcie / nvlink / nvlink2

Performance cost of writes under cudaMemAdvise(setReadMostly)

benchmark NCCL

full-duplex and half-cuplex unified memory

does cudaHostAllocPortable have a performance cost before UVA?

CUDA 3.0 3.2.5.1 describes portable as not the default, though no reason is given. Is there a performance cost?

full-duplex and half-duplex remote access

performance difference of mapping vs migration

observe bulk page migration in kepler

measure page fault cost

Linked-list traversal with large stride to avoid prefetching.

contention benchmarks

effect of cudaDeviceEnablePeerAccess on um coherence bandwidth

bandwidth GPU -> CPU wc memory

Linux DMA benchmark

full duplex and half-duplex bandwidth

CPU bandwidth writing to wc memory