TR-2016-09
ESP: A Statistical Approach to Predicting Application Interference
Nikita Mishra; John D Lafferty; Henry Hoffmann. 20 September, 2016.
Communicated by Henry Hoffmann.
Abstract
Independent applications co-scheduled on the same hardware
will interfere with one another. Predicting this interference
is a key step in scheduling independent applications
in large distributed systems. In this paper we
study several state-of-the-art regression models for estimating
application interference. We find that linear
models produce low-accuracy prediction, while interaction
(i.e., non-linear) models require so many input features
that their usage becomes impractical. We therefore
propose ESP, a hybrid regression model that uses linear
techniques to select a smaller set of input features,
and then uses non-linear techniques to estimate application
interference from these features. The result is a
highly accurate and practical regression model for application
interference estimation. To demonstrate this practicality,
we implement ESP and integrate it into a scheduler
for both single and multi-node Linux/x86 systems
and compare the scheduling performance to state-of-the-art
heuristics for scheduling to avoid interference. We
find that ESP-based schedulers are 1.25-1.8× faster than
these heuristics depending on the scheduling scenario.
Additionally, we find that ESP’s accurate interference
predictions allow schedulers to avoid catastrophic decisions
which heuristic approaches fundamentally cannot
detect.
Original Document
The original document is available in PDF (uploaded 20 September, 2016 by
Henry Hoffmann).
