Memory Stall Software Harvester

This repository contains the prototype of the Memory Stall Software Harvester (MSH), the first system designed to transparently and efficiently harvest memory-bound CPU stall cycles in software. Why harvest memory-bound stalls through a software mechanism when there are well-known hardware harvesting mechanisms like Intel Hyperthreads? The answer is that hardware mechanisms are inflexible: they cannot differentiate between latency-sensitive applications and others, and they only provide limited concurrency (e.g., 2 threads), often harvesting too much or too little. MSH allows for adjusting the length and frequency of cycle harvesting more precisely, providing a unique opportunity to utilize stalled cycles of latency-sensitive applications while meeting different latency SLOs. For more details about MSH, please take a look at our OSDI'24 paper.

Limitation in Prototype

Our prototype has some assumptions on primary and scavenger applications to simplify the implementation

• Primary application must use pthread library.
• To use a lightweight context switch, our prototype assumes that scavenger applications are given as shared objects and include several symbols and an entry point. Thus, you need to modify and recompile your scavenger application. The required transformation is straightforward and short.
• We assume x86-64 platform and use gcc/g++ with -fno-omit-frame-pointer and -mno-red-zone flags enabled.
  • Tested with gcc/g++ 11.4 and 22.04.1-Ubuntu (kernel version 6.5.0).

Basic Usage

Workflow

The use of MSH involves three pieces of software: profiler (scripts based on perf), binary instrumentation (llvm-bolt), and MSH runtime(libmsh). We assume that a user has a primary application and a set of scavenger applications that will run when there are stalled cycles in the primary application. One can use MSH in the following way.

• Modify scavenger applications
• Profile primary/scavenger applications.
• Modify primary/scavenger binaries through binary instrumentation with the profile result.
• Finally, run instrumented binaries with MSH runtime.

We'll show you how this workflow works with one simple primary(ptrchase) and scavenger(compute.so).

Prerequisite: Modify Scavenger

MSH requires a scavenger to have the following symbols in the file containing main function.

extern "C" {
int    crt_pos      = 0;
int    argc         = 0;
char **argv         = 0;
}

Then, rename main function to entry as shown below.

extern "C" int
entry(void) {
...
}

Lastly, compile the scavenger to a shared object file.

Prerequisite: BOLT

We implemented all the binary-level instrumentation in BOLT. Here is the patch instruction:

git clone https://github.com/llvm/llvm-project.git
git checkout 30c1f31
patch -p1 < msh_bolt.diff

Then, compile BOLT by following the instruction in BOLT page. We'll assume that llvm-bolt is in PATH from now on.

Profile and Instrument Primary

# Compile ptrchase
cd apps
mkdir build
make primary

# usage: ./do_prof_primary.sh [binary] [args]
# Pointer chase 50MB array
cd ${HOME}
./do_prof_primary.sh ./apps/build/ptrchase 13107200

Makefile in apps has rules that use BOLT to perform binary instrumentation. We assume that llvm-bolt is in PATH.

cd apps
make build/ptrchase.bolt

Profile and Instrument Scavenger

# Compile compute.so
cd apps
make scavenger

cd ${HOME}
echo "$(pwd)/apps/build/compute.so" > scav.txt
./do_prof_scavenger.sh

You can instrument scavenger in a similar way. However, you need to specify the average yield distance (in nanoseconds) in scavenger to bound it.

cd apps
YIELD_DISTANCE=100 make build/compute.so.bolt

Run with MSH Runtime

We used LD_PRELOAD trick to attach the runtime to the primary without recompilation.

cd libmsh
make all

cd ..
export LD_PRELOAD=$(pwd)/build/libmsh.so:${LD_PRELOAD}
export LD_LIBRARY_PATH=$(pwd)/build:${LD_LIBRARY_PATH}

echo "$(pwd)/apps/build/compute.so.bolt" > scav.txt
export MSH_SCAV_POOL_PATH=$(pwd)/scav.txt
export SKIP_FIRST_THREAD=1

./build/apps/ptrchase.bolt 13107200

Note: don't forget to reset LD_PRELOAD when you finish testing MSH. MSH runtime will intercept all the following pthread functions otherwise

TODO

• Add the use of more complex primary (e.g., tailbench) and scavenger (e.g., graph algorithm).
• Explain knobs

Developed by

Sam Son and Zhihong Luo