View all Seminars  |  Download ICal for this event

Performance Characterization and Optimizations of Traditional ML Applications

Series: M.Tech (Research) Colloquium

Speaker: Harsh Kumar

Date/Time: Nov 26 16:30:00

Location: Online Seminar - ON-LINE

Faculty Advisor: R. Govindarajan

In recent years, Deep Learning based methods have attracted a lot of attention and research – both from statistics and systems. These traditional algorithms are easily explainable and are pretty fast for smaller and medium-size datasets. However, in large organizations, massive datasets spanning a couple of million sample points are not rare. A lot of research has been done to design or adapt these traditional algorithms for such massive datasets. However, we find an apparent lack of a detailed systems-based study for these algorithms in the context of huge datasets.
In this work, we study the systems behavior and bottlenecks for these algorithms in the context of huge training datasets. As part of our work, we start with a performance characterization study, identify critical performance bottlenecks experienced by these applications, and then measure the reduction in performance stalls along with apparent benefits in terms of speedup after applying some of the well-known optimizations at the levels of caches, main memory, and computation. More specifically, we apply optimizations such as (i) software prefetching to improve cache performance and (ii) data layout and computation reordering optimizations to improve locality in DRAM accesses and show the performance benefits they can bring in these applications. Last, we evaluate the sensitivity of predictions and the improvement in performance when the computations on precise (float/double) inner variables are interpreted as relatively low-cost integer operations. These optimizations are implemented as modification on the well-known scikit-learn library.
We evaluate the impact of the proposed optimizations using a combination of simulation and execution on real system and performance measurement. Our optimizations result in performance benefits varying from 5% -- 27% on different ML applications.
This is an online colloquium. The teams meeting link for this is:

Speaker Bio:

Host Faculty: R. Govindarajan