[Colloquium] Reminder - Muhammad Santriaji Dissertation Defense/Jun 9, 2022

[Megan Woodward]

This is an announcement of Muhammad Santriaji's Dissertation Defense.
===============================================
Candidate: Muhammad Santriaji

Date: Thursday, June 09, 2022

Time:  1 pm CST


Location: JCL 298

Join Zoom Meeting
https://uchicago.zoom.us/j/91475498292?pwd=RHZFMEI3NEdxNE93dTdZSngvc2QvQT09<https://urldefense.com/v3/__https://uchicago.zoom.us/j/91475498292?pwd=RHZFMEI3NEdxNE93dTdZSngvc2QvQT09__;!!BpyFHLRN4TMTrA!4nOvOxS_EApHaKwo-EMLZh1Ha2nti4RmwP_w1y_xxaCoZS3P2BikxOu0klfqKKGf4r23-qAQBE9srlkTR2MdgjK7c0BAxpw$>

Meeting ID: 914 7549 8292
Passcode: 953175

Title: Incentivizing Flexibility and Cooperation in Computer Systems Using Feedback Mechanisms

Abstract: Many computer systems and applications, from small embedded systems to large datacenter have deployment requirements. Meeting this deployment requirement in a dynamic environment is challenging and requires flexibility from the application and the system.
Flexibility is the ability to trade-off the value of one measure space by adjusting the value of another measure space. For example, both DNN and approximate computing applications can reduce their runtime latency by sacrificing their output accuracy. However, managing this flexibility is difficult. Prior approaches do not incentivize flexibility and cooperation. In the single stakeholder scenario where applications come from one stakeholder, they do not cooperate with the application and system knobs which makes the deployment inefficient in terms of energy, output accuracy, and performance. In the multistakeholder scenario, they do not incentivize the flexibility of applications which make flexible application produce higher output error.
Our first contribution is GRAPE and MERLOT, a hardware feedback mechanism to meet latency requirements while reducing energy usage. GRAPE is a hardware control system for GPU that provides a soft guarantee to meet the performance requirements. Meanwhile, MERLOT is a real-time hardware scheduler that provides a hard real-time guarantee.
Our second contribution is ALERT, runtime management for Deep Neural Networks that decrease output error or energy usage while meeting latency requirements. ALERT cooperates the whole flexibility from both application and the system. ALERT uses a probabilistic feedback mechanism that predicts the energy, performance, and output accuracy of the applications during the runtime.
The third contribution is SIM, runtime management that incentivizes the flexibility of applications in the multistakeholder scenario. Prior approaches inadvertently disincentivize the flexibility by enforcing flexible applications to adapt to meet their deployment requirement, thus encouraging greedy behavior where every stakeholder deploys networks that consume as many resources as possible. SIM instead only enforces the adaptation to an application that holds the most resources. In each iteration, on behalf of the applications, SIM would make an application to a configuration that minimizes their output error such that the resource usage is either less than the application that holds most resources or higher as long as there are enough slack resources. SIM incentivize the deployment of flexible application by giving opportunity for all of the applications to fight for the slack resources by being flexible.

Advisors: Hank Hoffmann

Committee Members: Hank Hoffmann, Shan Lu, and Michael Maire


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.cs.uchicago.edu/pipermail/colloquium/attachments/20220609/cbe8d5eb/attachment-0001.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: Dissertation_Draft_Intro.pdf
Type: application/pdf
Size: 74800 bytes
Desc: Dissertation_Draft_Intro.pdf
URL: <http://mailman.cs.uchicago.edu/pipermail/colloquium/attachments/20220609/cbe8d5eb/attachment-0001.pdf>