Project Proposal Feedback

[pooyanjamshidi]

Problem Statement:

• First, great job addressing the initial feedback at the beginning of your submission. You've clearly outlined the problem around InfAdapter’s lack of energy consumption reporting and benchmarks. This is key for sustainable AI/ML.
• One suggestion: it would be stronger if you tie the problem to real-world challenges or examples to make it more relatable.

Solution:

• The use of etrace2 and Zeus to measure energy consumption is a solid approach, and I like the focus on improving accuracy, reducing latency, and minimizing energy.
• Here’s where you could take it further: I suggest you dive into the Deep RL for Scheduling and Resource Autoscaling papers we discussed with Misagh. Look through their references and pull out some solid work from top computer systems conferences, especially if they’ve released their code. This survey on Microservices might be useful too since they share similar challenges with ML pipelines.
• Also, check out papers on Deep RL for Resource Management that haven’t released their code, like SIMPPO: A Scalable and Incremental Online Learning Framework for Serverless Resource Management. Try reaching out to the authors to see if they might share their code—it’s worth a shot.
• Don’t forget about some of the classic papers on this topic, like this one from Mohammad Alizadeh at MIT: DeepRM GitHub repo. It’s in a different system context but still useful.

Evaluation:

• I’m glad to see your plan for benchmarking with different configurations. However, you might want to look at some of the top groups working on energy optimization, like Prashat’s recent work on carbon-first systems (he’s switched directions toward sustainability), and Abel’s work as well (he’s a friend, and they’re doing great stuff). Here are links to their papers:
  • Prashat’s works
  • Abel’s works
• Keep an eye out for papers using the term “carbon-first”—that’s the right direction for quality work in this area. For example, look at CarbonScaler: Leveraging Cloud Workload Elasticity for Optimizing Carbon-Efficiency out of UMass lab.

General:

• For future work, it’s a good idea to start identifying solid venues for submitting your work. Look at what Prashat’s group and his postdocs are publishing recently—they’re a good example of where to aim. A useful venue to keep in mind is ACM Energy. Start planning for publication early!

---

Project Management Suggestions:

• I know my comments can get long, and it might feel overwhelming, but try not to worry. Here are a few ways to manage everything:

1. Prioritize the evaluation setup: Your next big deliverable should be a clear presentation of one key challenge you’re addressing, backed up with experimental results. Think of it like a Jupyter notebook (so that an experiment would start and plot some results after a while, some examples in the IPA repo for replications. Think about preparing a plot that directly goes to your final report/paper), or at the very least some plots based on your own experiments offline (not other papers or fake data). This task is crucial to show you’re on the right track.

2. Use GitHub features: GitHub’s issues feature is great for project management. Break down my comments into tasks and assign them. Don’t forget to also track the tasks you identified in your proposal.

3. Coordinate with other IPA-ext team: You know Regan's team is extending IPA too and planning to do something related to energy, please consider collaborating with them. BUT be careful—pick just one task to collaborate on initially. Collaborating on too many things at once could be a recipe for disaster. Focus first, then expand collaboration if it works well.

Do your best to enjoy the process—it’s a big learning experience, not just about delivering a good project, but also about how you manage everything along the way. I want you to feel the satisfaction and the joy when you get to the finish line.