rivendile / muri

Artifacts for our SIGCOMM'22 paper Muri

0. Introduction

This repository contains the source code for our SIGCOMM'22 paper "Multi-Resource Interleaving for Deep Learning Training".

1. Content

• simulator/ contains code for simulation and is adapted from Tiresias. Please refer to <repo>/simulator/README.md for detailed information.
• cluster_exp/ contains code for real-cluster experiment. Please refer to <repo>/cluster_exp/README.md for detailed information.

2. Reproduce results (for SIGCOMM'22 artifact evaluation)

Please refer to <repo>/simulator/README.md and <repo>/cluster_exp/README.md for details.

Note: Due to the execution scripts of testbed experiments are highly related to intracompany platform, we only demonstrate the functionality and show the pseudocode of the related scripts (e.g., run.sh, prepare_env.sh). Please adjust to your platform if you would like to execute the testbed experiment.

3. Contact

For any question, please contact zhaoyh98 at pku dot edu dot cn

muri's Issues

A question about KVReader

There was an error like "ModuleNotFoundError: No module named 'dataloader'" in the file nlp_model.py(line 9:from dataloader import KVReader). I tried to use "pip install dataloader" to install the dataloader. But it didn't work. So I wonder if KVReader was written by the author?

Questions about the run_sim.py implementation

hello. While looking at the simulator code, I have a question.

While looking at the SRTF code (def shortest_first_sim_jobs(...)) in run_sim.py, I saw that the execution time including overhead was subtracted from the code that updates the remaining_iteration of the job in the runnable_jobs list.

In my opinion, this overhead is the time from the start of initial training to the start of iteration.
Therefore, it seems that this overhead needs to be subtracted only once at the beginning, but in this code, it seems that this time is repeatedly subtracted to calculate it.

Thanks for answering the question :)

a few questions about paper

Hello, I have a few questions about this paper.

I have sent you an email and look forward to your reply. Thank you!

Question about the implementation of Themis scheduler

I have a question about the themis simulator in the file run_sim.py, in the function themis_sim_jobs, it first call the function get_isolated_throughts to get a list of throughput of current runnable jobs when each of them runs alone with just 1 / N of the total resources

but in the function get_isolated_throughputs, it returns the varible allocation, not the varible isolated_throughputs in the function, I think it is a bug.

Hoping to your reply!