Midrange server is primed for running multithreaded apps
Not long ago, Sun announced the newest UltraSparc CPU, code-named Niagara, and officially dubbed the T1. What was readily apparent was that this newcomer wasn’t your daddy’s Sparc. Sun’s newly released Sun Fire T2000 UltraSparc T1-based server isn’t your daddy’s Sun Fire either.
Historically, a big horsepower Sun server such as the quad-CPU Sun Fire V440 is big in every way: physical size, power consumption, and price. Sun also began moving to lower-cost, lower-performance boxes such as the single CPU Netra X1. The T2000 is really the middle ground between the two, bringing higher processing power into a smaller chassis, but with an important twist.
The T2000 is technically a single-CPU system with its sole UltraSparc T1 processor. The difference is the T1’s architecture, which contains eight cores, each capable of handling four concurrent threads, for a total of 32 simultaneous threads. In fact, when pulling info on the CPUs from Solaris 10, 32 virtual CPUs are shown, each at roughly 1GHz. The UltraSparc T1 will be available in 1GHz and 1.2GHz flavors, with four, six, or eight cores.
Within the chip itself sits a 136GBps crossbar switch carrying data within the CPU, and 3MB of L2 cache that’s shared among the cores. Beyond this are four DDR2 channels, bringing memory bandwidth to 23GBps. Amazingly, Sun claims this CPU consumes less than 80 watts and runs considerably cooler than a comparable Xeon CPU, which runs more than 100 watts.
Although the UltraSparc T1 sports up to eight cores running at 1GHz or 1.2GHz, floating-point operations are farmed out to a shared FPU. This means that the T1 simply isn’t going to be great at heavy math, and certainly isn’t going to be the CPU of choice for non-threaded tasks, such as simulations.
What the T2000 and the T1 CPU are very good at is nearly any form of threaded tasks such as Web and application services. The T2000 is a 2U system pulling much less power than the 3U Opteron system I tested. Taking the results of my Web serving tests and extrapolating them against power consumption, rack density, and heat generation -- which directly links to datacenter cooling costs -- I found that the T2000 actually provides more bang for the buck.
Under the hood
The UltraSparc T1 is fully compatible with the Sparc V9 standard, so code written for any UltraSparc platform will run on the T2000. My evaluation T2000 came with a fresh install of Solaris 10.
Literally the first thing I noticed about the T2000 was its weight. For a 2U server with dual power supplies and four internal hot-swap drives, it’s surprisingly light. Its weight is no doubt attributable to the use of 2.5-inch SAS (serial attached SCSI) disks and the 2/3 rack-length chassis. The back panel houses the two hot-swappable power supplies, four embedded Gigabit Ethernet ports, a standard DIN9 RS-232 port, a separate system console port, two USB 2.0 ports, and an SC (subcriber connector) management Ethernet port providing lights-out system management.
There are also four expansion slots available: two PCI-X and two PCI-E. In my evaluation unit, one of the PCI-X ports was taken up with the SAS controller for the 2.5-inch internal drives, but Sun claims that this will be integrated onto the mainboard in upcoming production runs. Up front, the T2000 sports two more USB ports, the four hot-swap 2.5-inch drive bays housing 73GB 10K rpm SAS drives, and a DVD drive.
In the rack, the T2000 is loud but runs cool, being very adept at climate self-management and running the fans only when necessary. I racked the T2000 just above a 4P dual-core Opteron server, and the noise and power-draw difference was noticeable.
In testing the T2000, I ran some benchmarks to get a feel for how it drives. All in all, these results show the UltraSparc T1 and the Sun Fire T2000 as a solution to fit a specific need. Heavily threaded applications that are light on floating-point computations will perform well on the T2000, whereas less- or single-threaded applications will perform quite poorly when compared to Intel or AMD dual-core systems.
My first order of business was to compile and run Apache 2.0.54 on both the T2000 and on Red Hat Enterprise Linux 4 running on the quad dual-core Opteron. Here, there is no contest. The Opteron finished the configuration and building of Apache in approximately 2.12 seconds, which was less time than it took the T2000 to finish the configuration. These compilations were only utilizing a single thread and definitely highlight the Achilles’ heel of the T1.
Next, I ran 32 concurrent 4K DSA (Digital Signature Algorithm) crypto key generations on each server. Where the Opteron server took 44 seconds, the T2000 took 22 minutes. Based on this alone, it’s obvious the FPU performance of the UltraSparc T1 isn’t so great.
Next came the Apache Web serving benchmarks. This is where the T2000 held its own. Although the T2000 did not eclipse the quad-Opteron, it ran consistently at 85 to 90 percent of the performance of the Opteron system.
In real numbers, this meant that the quad dual-core Opteron server served approximately 2,500 pages per second of an 8KB static page with 10 image elements while the T2000 averaged 2,200 pages per second. A similar test on a dual 3.4GHz single-core Intel Xeon server running Windows 2003 and IIS was slightly lower than the T2000, at 2,025 pages per second.
My SSL tests showed the T2000 suffering slightly as compared to the dual-core Opteron, averaging 115 operations per second, with the Opteron server dishing out 379 requests per second. The dual-Xeon Windows 2003 server held the middle at roughly 270 ops per second.
In the T2000’s literature, Sun draws some financial comparisons to large Web frameworks built with Intel Xeon servers and then migrated to T2000 servers. One case study shows the cost reduction of a Web farm migrated from 1,000 dual-CPU Intel servers running Windows and WebSphere to 406 Sun T2000 servers running Solaris.
The claim is that the Sun Fire T2000s provided seven times the overall performance and nine times the performance per watt than the old farm, which is possible, depending on the clock speed of the old farm, reducing operating costs from $2.63 million to $550,000 per year. Even if the performance between the two farms was equal, that’s a big pile of money back in the budget.
The newest member of the Sun family might be considered a boutique server, but the UltraSparc T1 is truly a revolutionary processor. If it’s threaded performance you’re after, give the T2000 a look.
Overall Score (100%)
|Sun Fire T2000||8.0||8.0||9.0||8.0||9.0||8.0|
This weekend's Windows 10 upgrade has users angry, and it's unclear if the ploy will continue
Here’s the best of the best for Windows 10. Sometimes good things come in free packages
Speaking at the O'Reilly Fluent conference, Eich also endorsed the Service Workers mobile app...
Spoiler alert: There probably isn't. But that shouldn't cause anyone to panic aside from Wall Street...
Oracle says Java EE 8 will be equipped for cloud deployments, microservices, containers, and...
IoT will soon permeate every aspect of our lives -- the very definition of sprawl. How will we derive...
Git was made for distributed teams, but long distances introduce special challenges