HomePreparing for 227 CPUs: The Why and How of porting a new operating system to the IBM Blue Gene /L and /P supercomputers
Preparing for 227 CPUs: The Why and How of porting a new operating system to the IBM Blue Gene /L and /P supercomputers
Date:Thursday, April 23, 2009
Time:1:00pm - 2:00pm
Location: LBNLBldg. 50F, Room 1647
Speaker:
Dr. Ron Minnich
Sandia National Laboratories
Abstract:
Many predictions for the high end systems of 2018 envision computers
with 227 CPUs. The construction of such machines represents a break
with the progress of HPC: if we take a baseline to 1991, we can see
that machines at that time were being designed with 29 CPUs, and we
will see machines with 220 cores in 2012 (21 years, 11 bits in the
exponent). To meet the 2018 projections, the machines are going to
start to really grow: 1.2 bits per year, or double the historical rate.
What would an operating systems environment for 100M CPUs look
like? Can we expect that it will be Linux 2.18.200? Will it be a
souped-up Light Weight Kernel? Or none of these?
In 2002, a group of DOE researchers decided to answer some of these
questions. We saw a need to reinvigorate operating systems research in
the US research community. We started the FAST-OS working group,
which became the FAST-OS program in 2005. This program was recently
renewed for 3 more years.
I fell into the "none of the above" category. I felt that Linux had
severe deficiencies, which time has not improved; and that the Light
Weight Kernels did not do enough. In this talk I will describe our
work with Bell Labs and IBM in porting a new operating system to the
IBM Blue Gene/L and /P supercomputers. I am going to embed the port
talk in another talk, however, which provides more of the context for
this work. One might ask, "why another OS? Isn't everything just
running Linux?". I will answer that question with a quick review of my
2002 talks; and review my predictions of that time, and see how they
came out. I will then describe some of the work we have done with our
new operating system. If time and fate allow, I will even try for a
demo.
Host of Seminar:
Erich Strohmaier