Home arrow Get Informed arrow News arrow ITJungle: Sun Gangs Up Sparc T2+ Chips with Maramba Servers
ITJungle: Sun Gangs Up Sparc T2+ Chips with Maramba Servers PDF Print E-mail
Written by Timothy Prickett Morgan (ITJungle)   
Wednesday, 09 April 2008

There is still enough competition in the Unix space to make vendors hate to see each other get the limelight. Which is one of the reasons why Sun Microsystems has timed the launch of its "Victoria Falls" Sparc T2+ multicore processor and the related "Maramba" server line in the same week that IBM is launching big Power6 boxes and Hewlett-Packard is rolling out an update to its HP-UX Unix and a repackaging of the software.

The launch of the Victoria Falls processor and the two Maramba servers are basically what I told you they would be back in March. The Victoria Falls chips is a variant of the "Niagara-2" Sparc T2 processor that has had two of its memory controllers removed as well as the integrated "Neptune" 10 Gigabit Ethernet ports and replaced them with symmetric multiprocessing links that use some of the memory lanes on the processor to lash the two caches in the chips together into a two-way processor complex. The Niagara-2 chip runs at 1.2 and 1.4 GHz and has eight eight-threaded Sparc cores, and with the Victoria Falls design, Sun is able to put a two-socket Sparc T2+ server into the field, thereby making the Niagara designs more scalable and therefore more appropriate to the huge installed base of UltraSparc-II and UltraSparc-III iron sitting out there in the data centers of the world.

The talk out there in the blogosphere has made much of the fact that the Victoria Falls chip only has two memory controllers compared to the four memory controllers in the Niagara-2 chip, and has been critical of Sun's decision to remove the Neptune ports from the chip, too. The memory controller count is not an issue, explains Mat Keep, product manager for Niagara systems at Sun, and in fact, the Victoria Falls chip has sufficient memory bandwidth for the jobs it is intended for. Ditto for the integrated networking, which is not as necessary for the Maramba servers, which are aimed at database and application serving, as they were for the Niagara-2 class servers, code-named "Huron" and aimed at Web infrastructure workloads that were sitting on the edge of the network and hence where network performance is a competitive advantage.

As for the memory controller issue, Keep says that the original Niagara-2 chip design was going to use DDR2 main memory instead of Fully Buffered DIMM main memory because FB-DIMM memory was a lot more expensive, and because of the lower performance of DDR2 memory, Sun therefore planned to have twice as many memory controllers as you might expect. Then, at the last minute before Sun taped out the Niagara-2 chips, FB-DIMM memory prices dropped, and Sun switched to this technology while keeping the same number of memory controllers. The upshot is that the Niagara-2 chips delivered just under 60 GB/sec of memory bandwidth for a single socket machine, which is a lot of bandwidth. The Victoria Falls design delivers just 38 GB/sec of memory bandwidth per socket, or 76 GB/sec for the whole two-socket system, and that is not a big improvement. (The Niagara-2 chip had four controllers, but just two DIMMs per channel, compared to the Victoria Falls chip, which has two controllers but four DIMMs per channel.) Keep says that the Niagara-2 designs had way more bandwidth than the processors needed because of the design switch. The reality is, with the Sparc T2+ chip having the same pinout as the Sparc T2, you have to give up something if you want to create an SMP chip, and memory controllers and integrated networking is what Sun decided to give up. The market will decide if Sun did the right thing or not.

The market is probably going to be pleased, particularly customers with Sparc/Solaris applications who want to get off aging iron in the way that is least disruptive to applications. The jump from the Sparc T1 to the Sparc T2 chip was a factor of two in performance gain, clock for clock and core for core, and that was just because the chip went from 32 to 64 threads. (Obviously, that 2X performance is for software that likes threads; single-threaded jobs might see only a tiny performance boost, or none at all.) The Sparc T2 added a floating point unit for each core in the chip, as opposed to a single unit shared by all the cores, which helped performance on math-sensitive workloads. According to Keep, customers jumping from the Sparc T2 systems to Sparc T2+ systems will see anywhere from 1.8 times to more than 2 times the performance, and again, a lot of that is just because the thread count in the system is jumping from 64 to 128. The clock speeds on the T2+ chip are still at 1.2 GHz or 1.4 GHz, which is a bit disappointing. Sun should be able to get the clocks up in this chip to 2 GHz or higher, but is probably keeping speeds low to keep main memory and CPU cycle times closer together, thereby boosting the overall efficiency of the system. The other approach would be to use lots of L2 and L3 cache and crank clocks up to 3 GHz, but that burns more energy to add performance. (The Victoria Falls T2+ chip has 4 MB of L2 cache on the die, which is a relatively modest amount.) Energy goes up on a log scale as performance goes up on a linear--and relatively flat--scale.

The Maramba servers come in two flavors. The T5140 is a 1U rack-mounted server that can be equipped with T2+ chips running at 1.2 GHz with four, six, or eight cores activated per CPU socket. (The more cores activated, the higher the price.) This box has 16 memory slots, and supports from 8 GB to 64 GB of main memory using 1 GB, 2 GB, or 4 GB DIMMs. The system has room for up to four small form factor SAS drives in 73 GB or 146 GB capacities, and has four Gigabit Ethernet ports and three PCI-Express x8 peripherals on the board as well (two of them can be converted to support 10 Gigabit Ethernet XAUI connections). The entry configuration of this T5140 box comes with two T2+ processors with four cores activated (64 threads in total) and running at 1.2 GHz, plus 8 GB of memory and two 146 GB disks for $14,995. A machine with six active cores per T2+ chip (96 threads) and 16 GB of memory costs $16,995, and a machine with the full 128 threads turned on in the box plus 32 GB of memory costs $29,995.

The Maramba T5240 is a 2U rack-mounted server that has room for up to 16 2.5-inch SAS drives and a DVD player. It sports two T2+ chips linked together with four, six, or eight cores running at 1.2 GHz or eight cores running at 1.4 GHz. The machine has 32 memory slots, and main memory can scale from 8 GB to 128 GB. The machine has the same four Gigabit Ethernet ports on the board as the T5140 server, but it has more expansion slots, with four six PCI-Express x8 slots (two again convertible to 10 Gigabit Ethernet XAUI connections). The base T5240 costs $17,995, and it has two six-core T2+ chips (96 threads) running at 1.2 GHz, 8 GB of main memory, and two 146 GB disks. Activating all the cores running at 1.2 GHz and increasing main memory to 32 GB drives the price up to $31,995 with two disks. Sun did not provide pricing for a machine with the 1.4 GHz cores on its Web site.

Both Maramba servers come with Solaris 10 8/07 (Update 4) preinstalled on the boxes; a whole bunch of Sun middleware and development tools are also preinstalled. While the initial Niagara-1 machines supported Ubuntu 7.04, this Linux variant is not yet available on the Sparc T2 or T2+ machines. The Linux kernel has had support for the T1 architecture since April 2006, so it is just a matter of the Linux distros adding and testing this code for their implementations of Linux. Gentoo (a Debian variant like Ubuntu) and CentOS (a clone of Red Hat Enterprise Linux) have both just completed work on support for the T1 chips and are moving ahead to the T2 and T2+ chips, according to Keep, and the OpenBSD Unix project has just finished up its T1 support, too.

Next week, we'll talk about how Victoria Falls is just the beginning of a more scalable Niagara product line.

 

Read the original article: http://www.itjungle.com/bns/bns040908-story03.html

Comments (1)add comment
EdwardOCalaghan: Gentoo is *not* a Debian variant
Gentoo is *not* a Debian variant, btw.
Its a source based distro that no one uses :p
Regards
1

report abuse
vote down
vote up
May 01, 2008
Votes: +0

Write comment
quote
bold
italicize
underline
strike
url
image
quote
quote
smile
wink
laugh
grin
angry
sad
shocked
cool
tongue
kiss
cry
smaller | bigger

security image
Write the displayed characters


busy
 
< Prev   Next >
impersonal-mites
Generated in 0.532347202301 Seconds