I have to admit that I get a little excited whenever a new generation of scorch-your-eyebrows-off CPUs hits the server market. So when I got a chance to check out the newest Intel Nehalem Xeon systems, I made sure my insurance policy was up to date and relocated any breakables to another room.
Intel's newest server processors represent a major architectural change from earlier Xeon generations. One big improvement over previous-generation Xeons is the addition of an onboard memory controller and Non-Uniform Memory Access (NUMA). Taking a cue from AMD, Intel added NUMA to Nehalem to help eliminate cache starvation by tying banks of RAM to each processor.
Intel also redesigned the I/O system between CPU and peripherals, doing away with the front-side bus bottleneck and replacing it with a high-speed pathway called QPI - Quick Path Interconnect. QPI can transfer data to local peripherals as fast as 25.6GBps, nearly double the performance of a 1600MHz front-side bus system.
Other improvements include the reintegration of Hyper-Threading, the elimination of the Northbridge controller (PCI Express and Direct Media Interface are now in the CPU) and support for DDR3 memory.
Enter the dragon
I was able to secure three tower servers based on Intel's latest Xeon offering from Dell, Fujitsu, and Hewlett-Packard, and from the moment I fired each one up for the first time, I knew these servers were something special. The best part was watching Windows Server 2008 (my requested operating system) boot up in about 52 seconds — simply fantastic.
In order to make this an apples-to-apples comparison, I requested a similar configuration from all three server vendors. (IBM was also invited, but did not provide a system.) Each server came with two Xeon X5550 2.66GHz CPUs, 24GB of 1,333MHz DDR3 RAM, and Windows Server 2008 Standard Edition (64-bit). I asked each vendor to supply hard drives in a RAID 5 configuration but left the size and type of drives up to each vendor. Dual Gigabit Ethernet and multiple USB 2.0 ports were standard on each system.
To get a feel for each platform's performance, I ran the SPECjbb2005 Java server benchmark and the STREAM memory benchmark tools on each system. The SPECjbb2005 test is a Java virtual machine stress test that emulates a three-tier client/server order and inventory system. During the test, a number of virtual warehouses is created (two warehouses per processor core), with each warehouse handling order entry, payment, status, delivery, and reporting transactions. Additionally, SPECjbb2005 also measures the performance of the CPUs, caches, memory, and the scalability of shared memory.
Rock around the clock
Using SPECjbb2005, I was able to make all 16 cores (8 Hyper-Threaded cores) busy, achieving near 100 percent utilization at the end of the test run. The test suite increased the workload by one until it reached the maximum of 32 virtual warehouses (2 warehouses per core). At the end of the test run, SPECjbb2005 generates a score in bops (business operations per second, and SPECjbb2005's unit of measure) for each warehouse simulation. Bops represent the overall throughput achieved by all the warehouses in a test run; a higher bops value indicates better overall performance.
Test Center Scorecard
|
Performance |
Expandability |
Management |
Power Usage |
Value |
Overall score | |
| 40% | 20% | 20% | 10% | 10% | ||
|
Dell PowerEdge T610 |
10 | 7 | 9 | 9 | 8 | 8.9 (very good) |
| Fujitsu Primergy TX300 | 10 | 9 | 9 | 7 | 8 | 9.1 (excellent) |
| HP ProLiant ML350 | 10 | 8 | 9 | 8 | 9 | 9.1 (excellent) |
The SPECjbb2005 results of all three servers — 219,342, 218,527, and 217,820 bops — varied by less than 1 percent, showing performance of the systems to be essentially identical. This is due in large part to the similarities in the underlying architecture of the motherboards. For comparison with a previous server generation, I ran SPECjbb2005 against an HP ProLiant ML360 G3 and saw just how far the technology has come. My G3 ProLiant has a pair of 2.8GHz dual-core Xeons (533MHz FSB, 512MB of L2 cache) and 4GB of PC2100 DDR RAM. SPECjbb2005 was only able to run up to eight warehouses (my dual-core Xeons were not Hyper-Threaded), but the results were only about 5 percent of what the new Nehalem-based systems posted. For example, at eight warehouses, the three Nehalem servers averaged 153,259 bops, while my G3 ProLiant scored only 7,596 bops — not even close.
While the load testing took place, I monitored power usage using the Watts Up Pro 'plug load' meter. Watts Up Pro logged all of the collected data via USB cable to a client PC where I was able to view watts, voltage, amperage, and other metrics specific to power consumption.
This is one area where I did see a variance among the servers. All three servers drew very little power while in standby mode (plugged in but powered off). Once the systems were powered on and booted up, the Dell PowerEdge consumed fewer watts under load than the HP server and substantially less than the Fujitsu server. The potential impact on your power bill? Assuming a rate of 14 cents per kilowatt hour (roughly the US average), the Dell would cost about $100 less than the Fujitsu, and about $25 less than the HP, to run 24 hours per day for a year.
The STREAM 5.8 memory benchmark allowed me to see how fast bytes could be moved in and out of memory. STREAM measures sustainable memory bandwidth in megabytes per second, and all three servers recorded performance numbers that back up Nehalem's architectural advances. As with SPECjbb2005, the STREAM Triad results — 33,463, 33,508, and 33,536 — were again very close together, varying less than 1 percent across the three servers.
Server virtualization is high on most admins' checklists, and the Nehalem Xeons are built to provide support for all popular virtualization packages. All three servers are identical in this regard, supporting bare-metal hypervisors as well as OS-based virtualization without a problem.
Next: Rankings
Join the CIO Australia group on LinkedIn. The group is open to CIOs, IT Directors, COOs, CTOs and senior IT managers.