|
Last week, Sun Microsystems
began shipments of its first two-socket variants of its "Niagara"
family of servers, marking a new phase in scalability for its multicore
Sparc server product line. While Sun still sells plenty of big iron,
much of it is too big and too expensive for the huge installed base of
customers using machinery based on UltraSparc-II and UltraSparc-III
processors, which date from the late 1990s and the early 2000s,
respectively. The answer, of course, is to make more scalable Niagara
servers using the Sparc T family of chips.
Last week, Sun Microsystems
began shipments of its first two-socket variants of its "Niagara"
family of servers, marking a new phase in scalability for its multicore
Sparc server product line. While Sun still sells plenty of big iron,
much of it is too big and too expensive for the huge installed base of
customers using machinery based on UltraSparc-II and UltraSparc-III
processors, which date from the late 1990s and the early 2000s,
respectively. The answer, of course, is to make more scalable Niagara
servers using the Sparc T family of chips.
Tongues
have been wagging for years that this has, in fact, been Sun's plan all
along: to take the T1 chip, goose it with more threads and floating
point units, and then mesh multiple chips together into a family of
symmetric multiprocessing (SMP) servers that could compete with various
X64-based SMP boxes and do the one thing that X64 boxes--even those
running the X64 variant of Solaris 10--cannot easily do: run
application code that was designed specifically for the Sparc
architecture.
Being a
cut-down and threaded variant of the UltraSparc-II, the Niagara family
of chips, which include the original "Niagara" Sparc T1, the
"Niagara-2" Sparc T2, and now the "Victoria Falls" Sparc T2+, are
perfectly suited to running applications that are currently on aging
Sparc iron. And while the dual-core UltraSparc-IV+ machines made by Sun
and the dual-core Sparc64-VI machines made by Fujitsu-Siemens
and resold since last year by both companies as the Sparc Enterprise
line have lots of bandwidth and oomph, they are not cheap, they are not
cool (in terms of temperature), and they are not small. The Niagara
family of machines offers better bang for the buck on many workloads
and better performance per watt, particularly with the new T2+ machines
(I am extrapolating on that one).
This
explains why the Niagara class of servers have grown to a $1.3 billion
annualized run rate after only a year and of half of sales, a figure
that was provided to me by Mat Keep, product manager for Niagara
Systems at Sun, during last week's Victoria Falls Sparc T2+ chip and
its "Maramba" T5140 and T5240 servers.
The
Maramba server put two T2-style processors in a single 1U or 2U server
and they are hooked together through their memory buses to present a
single system image to main memory and the operating systems that run
on that image. Last week, Keep confirmed that Sun was indeed working on
a four-socket box that would use another variant of the Victoria Falls
chip, which we have heard is code-named "Botaka" from a number of
sources and which, if history is any guide, will be a 4U chassis and
will be called the T5440. The Maramba platforms got their name from one
of the rivers that feeds Victoria Falls, which separates the African
countries of Zambia and Zimbabwe, and Botaka is the name of one of the
gorges off the falls. I joked back in March that if you wanted to make
a big-dog server out of lots of T2-class chips, you would name it
Zambezi, after the main river that feeds the falls in Africa. As it
turns out, Zambezi is indeed a code-name at Sun, but it is the
code-name for the crossbar switch electronics that will be used to glue
four T2+ chips together into a single system image in the T5440 server,
which Keep says Sun will deliver sometime in the second half of 2008.
Now,
here's the interesting bit. The two T2+ sockets in a Maramba server
link together gluelessly--meaning that they have the electronics right
on the chip to link their memories together and you do not have to do
any more. (This extra circuitry for SMP clustering is the main reason
why Sun cut out the "Neptune" 10 Gigabit Ethernet ports that were part
of the T2 chip when it moved to the T2+ design.) But once you move to a
crossbar switch architecture, then there is no reason to stop at four
sockets. If the crossbar is designed correctly, of course.
Sun
is not saying that eight socket or larger Niagara machines are in the
works. "There are more multi-socket versions coming," Keep said, but as
for a machine ganging up lots of T2+ chips, "there is no plan to do
that right now" is all he would say.
True
and not true. The UltraSparc RK chip, according to the documents I saw
years ago, are comprised of 16 Sparc cores, divided into groups of four
that share L1 caches that are in turn linked to each other and to four
banks of L2 cache by a crossbar switch embedded on the chip. Sun
clearly has a very scalable crossbar inside the Rock chip, and it is
also likely that this crossbar can be extended to support multiple Rock
chips in a single system image, perhaps scaling from two, four, or
eight sockets, each with 16 cores inside each socket. Whatever Sun can
do for Rock, it can do for the T2 chip.
And
I am beginning to think, once again, that the differences between
Niagara and Rock chips are not as great as we have been led to believe.
To be sure, Rock will implement transactional memory and scout threads,
which Niagara chips appear to not do. But if those new technologies do
not pan out--and remember, the Rock chips were just pushed out a year
to the second half of 2009--do not be at all surprised to see a goosed
up Niagara chip suddenly plunked down inside the Supernova server
frames. If something is working, you tend to stick with it. Sun ditched
the "Millennium" UltraSparc-V designs in 2004 because they were not
going to pan out, and switched to goosing the UltraSparc-III by adding
cache and making it multicore, and hence the UltraSparc-IV. This would
not be the first time such a thing has happened.
Read the original article: http://www.itjungle.com/tug/tug041708-story01.html
|