Sequent stretches SMP limits
As chairman and chief executive officer of Sequent Computer Systems Inc.,
Casey Powell is pushing Intel processors into high-end markets where computers have to be
big and fast. In his last job, Powell was general manager of microprocessor operations at
Sequent's distributed shared-memory servers are opening the door for
nonuniform memory access (NUMA) technology based on Intel's four-processor chip sets.
Sequent of Beaverton, Ore., supplies computers to the Veterans Affairs Department,
Internal Revenue Service and Army.
GCN senior editor Florence Olsen interviewed Powell during a recent trip to
GCN: Why has NUMA gotten so hot?
POWELL: With symmetric multiprocessing (SMP) systems, we've seen the
same fundamental architecture for the last 14 years. We've come so far that we're now
running large data centers on this architecture. The problem is, we haven't been able to
make SMP systems big enough. Massively parallel processing (MPP) machines are extremely
hard to use.
NUMA lets us combine the best of SMP and the best of MPP. We really don't know of any
architectural limit to the number of processors we can hook together in the NUMA
architecture, and you need only one instance of the operating system.
Each node in an MPP machine has its own operating system and its own local memory, so
you treat it as a bunch of loosely coupled computers. That becomes a programmer's
For users trying to interact with the machine, it's almost impossible. MPPs are great
if you want to program to calculate the effects of a nuclear winter, but not so good for
NUMA combines all the benefits of SMP that we've been seeing since 1984. It's very easy
to use. The first level of trickery fools the user into thinking there's only one
processor in the machine.
Next the architecture tricks the operating system into thinking it has one large bus
when that's not the reality at all.
GCN: Why are these systems so expensive if they are based on standard
POWELL: We do not run the standard Intel quad-processor system boards
in our NUMA servers, but we run the standard Intel quad chip set. Instead of designing the
quad boards to 99.9 percent uptime, we design them to 99.99 percent uptime. That's for a
very good reason.
Statistically, if you put together 300 quads that have 99.9 percent uptime, you get 80
percent uptime. We need extremely high reliability and availability in the market we're
going after. Our implementation uses more layers in the backplane and more expensive
GCN: Are there many applications written for your Dynix PTX operating
POWELL: Dynix PTX is a standard Unix operating environment. There are
as many apps for Dynix as there are for Unix. Some may have to be certified that aren't,
but it's not as if we have a proprietary operating system.
GCN: Why did you choose gallium arsenide for your data pump?
POWELL: It lets us build very big, complex systems. Our data pump
running at 1 gigabyte/sec flushes data at a horrendously high speed. After you've done the
initial off-system fetches, the machine starts to stabilize.
Most applications with which we have experience show a 99 percent cache-hit rate. Other
interconnect schemes are much less complex, but of course the system performance and
scalability are lower.