search

osalbahr / pmem-redirection

I/O Redirection of Workflows to Local and Disaggregate Persistent Memory. This repo is for CSC 499, Summer 2023.
https://arcb.csc.ncsu.edu/~mueller/cluster/arc/
GNU Affero General Public License v3.0
1 stars 1 forks source link
optimization performance pmem

readme

PMEM Redirection

Goal: I/O Redirection of Workflows to Local and Disaggregate Persistent Memory

Installation

If you have root access, you might want to use your favorite package manager for these libraries/packages, with the exception of maybe installing libpmem2 from source if it's not offered for your system.

This project was built on NCSU ARC Cluster, which uses CentOS Linux 7 (Core) x86_64. You might need to add some needed dependancies that don't exist on your system. Try cat fetch.txt for more system information.

Memkind Library

See memkind#building-and-installing.

Because I don't have root permissions, I am installing in $HOME/.local.

./autogen.sh
./configure --prefix=$HOME/.local
make
make install

You'd also want to add the following to your .bashrc:

# man memkind
export MANPATH=$MANPATH:$HOME/.local/share/man

# For some reason, this is needed in addition to `-L`
export LD_LIBRARY_PATH=$HOME/.local/lib

# TODO: maybe do the same for #include and ld's lib, instead of `-I` and `-L` respectively

You can choose to install it at a system level, or to another directory. If you install at a system level, you probably don't need to modify the scripts to account for this change. If you're curious to see where .local is hardcoded, you can run the following command:

grep -r '[.]local'

Persistent Memory Development Kit (PMDK)

The Persistent Memory Development Kit (PMDK) is a growing collection of libraries and tools. Tuned and validated on both Linux and Windows, the libraries build on the DAX (Direct Access) feature of those operating systems which allows applications to access persistent memory as memory-mapped files, as described in the SNIA NVM Programming Model.

See Installing PMDK. I went with Installing PMDK using Linux Packages. If you need libpmem2, you should use Installing PMDK from Source on Linux.

Debugging

If you have root access, you might want to use your favorite package manager for these too.

I was debating using Homebrew since it can technically streamline installing from source, even without initally having root access, but is not officially supported (and faced some issues). I also just wanted to try out installing from source more anyways.

ltrace

For debugging a "realloc error", I wanted to have ltrace as follows. See their INSTALL for instructions.

git clone https://gitlab.com/cespedes/ltrace.git
cd ltrace
./autogen.sh
./configure --prefix=$HOME/.local
make
make install

# You can also use
./autogen.sh && ./configure --prefix=$HOME/.local && make && make install

Valgrind

To further investigate the memory error, I wanted to have valgrind. See https://valgrind.org/downloads

curl -O https://sourceware.org/pub/valgrind/valgrind-3.21.0.tar.bz2
cd valgrind-3.21.0
./configure --prefix=$HOME/.local
make
make install

# You can also use
curl -O https://sourceware.org/pub/valgrind/valgrind-3.21.0.tar.bz2 && cd valgrind-3.21.0 && ./configure --prefix=$HOME/.local && make && make install

GEF (GDB Enhanced Features)

Since gdb rpm's are outdated, I ended up needing to spack install gdb && spack load gdb. I initially installed gdb from source, but had glibc version issues and thought I'd use a tool that automates the whole process, spack.

Benchmarks

I experimented with fio, a well-known tool for I/O benchmarking. Here are some simple benchmarks with the following target files:

  1. $HOME/fio.txt - Network Filesystem (NFS)
  2. /mnt/fsdax/$USER/fio.txt - Persistent Memory (PMEM)
  3. /tmp/$USER/fio.txt - SSD (tmpfs)
[osalbahr@c63 fio]$ head *fio
==> nfs-home.fio <==
[nfs-home]
rw=randrw
size=${SIZE}
filename=${HOME}/fio.txt

==> pmem-fsdax.fio <==
[pmem-fsdax]
rw=randrw
size=${SIZE}
filename=/mnt/fsdax/${USER}/fio.txt

==> ssd-tmp.fio <==
[ssd-tmp]
rw=randrw
size=${SIZE}
filename=/tmp/${USER}/fio.txt

Here is the interesting part after running the default random read/write benchmarks with 1GiB of memory (full log below):

[osalbahr@c63 fio]$ SIZE=1GiB fio *.fio
nfs-home: (g=0): rw=randrw, bs=(R) 4096B-4096B, (W) 4096B-4096B, (T) 4096B-4096B, ioengine=psync, iodepth=1
pmem-fsdax: (g=1): rw=randrw, bs=(R) 4096B-4096B, (W) 4096B-4096B, (T) 4096B-4096B, ioengine=psync, iodepth=1
ssd-tmp: (g=2): rw=randrw, bs=(R) 4096B-4096B, (W) 4096B-4096B, (T) 4096B-4096B, ioengine=psync, iodepth=1
...
Run status group 0 (all jobs):
   READ: bw=12.1MiB/s (12.7MB/s), 12.1MiB/s-12.1MiB/s (12.7MB/s-12.7MB/s), io=477MiB (500MB), run=39364-39364msec
  WRITE: bw=12.1MiB/s (12.7MB/s), 12.1MiB/s-12.1MiB/s (12.7MB/s-12.7MB/s), io=477MiB (500MB), run=39364-39364msec

Run status group 1 (all jobs):
   READ: bw=491MiB/s (515MB/s), 491MiB/s-491MiB/s (515MB/s-515MB/s), io=476MiB (499MB), run=969-969msec
  WRITE: bw=493MiB/s (517MB/s), 493MiB/s-493MiB/s (517MB/s-517MB/s), io=478MiB (501MB), run=969-969msec

Run status group 2 (all jobs):
   READ: bw=25.1MiB/s (26.3MB/s), 25.1MiB/s-25.1MiB/s (26.3MB/s-26.3MB/s), io=477MiB (500MB), run=19036-19036msec
  WRITE: bw=25.0MiB/s (26.2MB/s), 25.0MiB/s-25.0MiB/s (26.2MB/s-26.2MB/s), io=477MiB (500MB), run=19036-19036msec
...

For the full log, see this gist.