|
30 minute interview
“We are writing history here; this is an inflection point”
 |
David W Yen, PhD Executive Vice-president
Scalable systems
Sun Microsystems |
There has been considerable interest in what Sun is doing
ever since you scrapped US-IV and announced Niagara and Rock. What’s the
latest on that front?
Niagara is a processor that will be deployed in both boxes and blades. It will
be targeted at commercial set-ups and telcos. The processor will have 8 cores
that will run 4 threads each to offer 32 threads in all. The Rock processor
design will offer the best of both worlds. Additional innovations around throughput
and processor core architecture, as well as Java processing, will be there in
the Rock chip, while single-thread performance remains very competitive. It
will be out by 2007-08. We are writing history here. From a technical perspective,
this is an inflection point.
In the Rock design, transistors are devoted to processing. It is meant for raw
computing with a Niagara derivative as a companion chip to handle other tasks.
In the case of Niagara, the design supports computing and networking interfaces.
Down the line, we plan to support cryptology in Niagara with public keys stored
in the chip.
There was time when MHz (later GHz) was the only measure
of processor performance. That’s changed even on the desktop. What lies
beyond raw gigahertz?
Speed or MHz is only one parameter through which you can measure processor performance.
In real business environments, more GHz offers diminishing returns. What is
important is how you ensure that the processor is 100 percent productive. Take
the Alpha processor which sits idle 70 percent of the time and it’s clear
that you have to feed the hungry processor.
All systems will have to employ memory interleaving. From the processor to the
memory, everybody uses the same DRAM. Special attention must be given to ensure
that new sub-systems match the processor step for step. My engineers are working
with the Solaris and JVM/JES (Java Virtual Machine/Java Enterprise System) guys.
When you start working on multi-core, multi-threaded systems, the question that
comes up is—how many threads? This decision has to be related to memory
latency. Then you have to figure out which market you are aiming at—data
centres or Web services?
 Many
years ago Sun had a chip called MAJC. Has this project influenced the company’s
current processor designs?
The current architect for Rock was the architect for MAJC.
That chip started us thinking about multi-core, multi-threaded architecture.
Some ideas from there are being applied to Rock.
While coding application software, performance improvements
are obtainable by writing directly to the OS. Is something similar possible
by writing directly to the hardware?
We have to be careful as Solaris-SPARC and Solaris-x86 share the same source
tree. In process implementation, we try to match the integration of our application
software with Solaris, and for network intensive operations we can provide a
hardware assist. For the TCP/IP stack in Solaris on Niagara we will have extra
hardware. Instead of doing an operation in the software, we will do it in the
hardware. [Parts of] Solaris will run in Niagara [processor’s microcode].
To run Java well, you need a good implementation of threads.
We have figured out a special recipe for Rock; it will run Java faster with
more parallel garbage collection. We will work to resolve the locking mechanism
that usually becomes a performance bottleneck.
eBay is a major customer for Sun. Does the experience of
handling that kind of workload help you design processors that satisfy the requirements
of your other customers?
The eBay auction workload is what I will call a ‘dotcom’
workload. A ‘business’ workload shares many of the characteristics
of a ‘dotcom’ workload. In both cases, the compute demand for each
thread is relatively light, but the volumes are huge. Every user who logs in
creates a thread.
There are multiple instances, hundreds of thousands of them—the Niagara
family of processors is optimised for that kind of workload. Rock is for the
HPC (high performance computing) workload with lots of floating point operations.
It will be a more traditional processor design when compared to Niagara, which
will be closer to a system-on-a-chip.
You have a large development centre in India. Is any hardware
work being done here?
10 percent of Sun India Engineering Centre’s 700 engineers are in my group.
They are working on chip design, CAD and quality assurance work for new products.
We have some people working on mainframe migration software. India and China
are large territories both in terms of market and talent pool.
TI is your fab partner. How do you characterise the relationship?
TI has been a good partner. Our relationship goes beyond financial transactions.
They are fabricating the UltraSPARC IV and IIIi. They are going to be the foundry
for Niagara and Rock.
| In 2005 or 2006, Sun plans to release processors
that provide 15x throughput of existing ones. After that, it wants to release
systems with 30 times today’s performance. The goal is simple, if audacious—a
single blade shelf that does the work of 32 of today’s 4-way servers;
eight rack units instead of 160; less than 3 kilowatts of power versus 38;
one blade system to manage instead of 32 servers.
Moore’s Law is slowing. CPU clock frequency that
doubled every two years may double in ever-longer increments of time.
Memory speeds are doubling every six years, creating a gap in the performance
of chips and memory. The end-result is that today’s processors are
sitting idle as often as 75 percent of the time, waiting to fetch data
from the system memory.
Sun’s approach to solve this knotty problem is to
use the additional transistors afforded by Moore’s Law to put multiple
cores on a single piece of silicon, with each core processing multiple
threads simultaneously. When one thread is waiting for memory, the affected
core will start processing another thread, thereby side-stepping the memory
latency issue. This approach is expected to improve chip efficiency and
thereby boost application performance. Will it work? We’ll have to
wait and see, but if it does, Sun would have pulled a rabbit out of its
R&D hat. The company succeeded in doing that before, and there’s
no telling if it might do it again.
|
—Prashant L Rao
|
