Date of Event:
June 14, 2016Where:
TIG Corporate Headquarters
10240 Flanders Court, San Diego, CA
Case Study: Predicting Compute Memory Needed for a Computer Job, Tuesday, June 14, 2016
6:00 PM – Networking
6:30pm -- Recap of Hackathorn, Gina Escobar
7:15pm -- Qualcomm Case Study
8:00pm -- close
Recommender System for a Compute Memory Request for a Computer Job by
Taraneh Taghavi, Qualcomm and Maria Lupetini, Qualcomm
Billions of computer jobs are run on Qualcomm compute clusters in a year to design and test semiconductor chips. There is a need to use all compute resources optimally, i.e. processors, storage and memory.
Currently, the engineers estimate the compute memory needed to run a job. When compared to what is actually needed for execution, their estimates overall are too high. This results in too much compute memory being allocated to these computer runs, and thereby reducing the ability to run other jobs on the same multi-processor server. This reduction in throughput delays precious design and testing time. In addition, this behavior drives compute memory purchase costs higher than necessary.
This presentation will highlight the data exploration, development, and deployment of a recommender system that predicts the amount of memory a compute job will need. We will discuss the array of statistical and machine learning techniques we explored, and how we selected the final model.