Computer Science Thesis Proposal
— 11:30am
Location:
In Person and Virtual - ET
-
Traffic21 Classroom, Gates Hillman 6501
Speaker:
JALANI WILLIAMS
,
Ph.D. Student, Computer Science Department, Carnegie Mellon University
https://www.cs.cmu.edu/~jalaniw/
Setup Times in Multiserver Systems
Many modern multiserver systems (e.g. datacenters) dynamically turn servers on and off in response to variation in service demand. Performing "dynamic scaling" in this way allows a system provisioner to save on operating costs during periods of relative downtime. However, this improvement in operating efficiency is not without its own cost. Since each server requires a "setup time" before it is ready to serve jobs, the delay performance of a dynamically-scaled system must necessarily be worse when compared to an equivalent system which always has all of its servers on. The question we are interested in is: how much worse? More completely, "How do setup times affect the delay performance of modern multiserver systems?"
Although dynamic scaling policies are quite popular in practice, theoretically-speaking we understand very little about how setup times affect the performance of real systems. Historically, research on dynamic scaling has taken one of two approaches when modeling the effect of setup times. Existing work either 1) ignores the effect of setup times entirely or 2) assumes that setup times are Exponentially-distributed in a bid to simplify the system's dynamics.
In my thesis, I demonstrate that both of these approaches tend to severely underestimate the negative effect that setup times can have on delay in modern multiserver systems. In reality, setup times cannot be ignored; setup times can cause profound increases in delay, especially when setup times are of a fixed length. I show this by giving the first-ever theoretical analysis of the delay in a multiserver system with Deterministic setup times, proving multiplicatively-tight upper and lower bounds on the mean delay for a queueing system which is a natural extension to the M/M/k queue. These results are made possible via a novel approach to steady-state analysis which might be of independent interest.
Thesis Committee:
Weina Wang (Chair)
Mor Harchol-Balter
Alan Scheller-Wolf
Jamol Pender (Cornell University)
Bill Massey (Princeton University)