Dual-processor, pizza-box servers have been a staple of computing infrastructure for much of the last 10 years. When an IT organization says it's building a new cluster or adding capacity to a server farm, it is invariably referring to purchasing these commodity servers, sliding them into racks, tying them into the network, and firing them up. In this sense, these systems are truly commodity items — the individual bricks that make up the computing edifice. To get an idea of the current state of the art, we approached three vendors (Dell, Hewlett-Packard, and Lenovo) to let us examine their standard, two-processor servers and see what they offer and at what cost.
Predictably, we found these machines were a lot alike in most ways that count to IT: processing power on a dollar-adjusted basis was similar. The areas of difference tended to be small features, the importance of which depends on specific needs of each site. The good news is that all three vendors delivered good products at favorable prices. Choosing any one of the models we reviewed will produce a satisfactory result. Consequently, we expect most organizations to continue using the same brand of vendor they have already settled on. However, the details of these models might well induce you to examine more of your vendor's server configurations.
The journeyman server
Not all pizza-box servers are the same. The journeyman category we're examining is the middle segment of a continuum bordered on the low end by value-oriented servers used by SMBs, and on the top end by heftier models, such as the virtualization-oriented systems I examined in a previous review.
Virtualization servers are designed to host many virtual machines. As a result, they typically have higher RAM and processor requirements. In addition, they tend to devote less space to on-board I/O, as the data for the VMs is almost always located in a spindle farm and piped to the virtual machines. This design enables virtual machines to migrate easily among hardware platforms — a principal value of virtualization.
While the virtualization servers generally sport four processor sockets to provide the computing heft, the journeyman servers are mostly two-processor affairs. This is because a dual-processor system is significantly less expensive, and it is more easily replaced in the event of failure. In other words, the two-CPU boxes are cheaper, and by running fewer tasks they represent a less disruptive point of failure. The preference for dual-processor systems also has drawbacks of course, not least of which is the additional complexity and cost that each incremental system imposes. These costs include software licenses, ports on the net, and the like.
In terms of I/O, journeyman servers tend to have a large complement of disk drive bays on board. They do not adopt the virtualization server's approach of diminishing local I/O devices. For example, the Dell and HP models have bays for 8 hard disks, while the Lenovo has room for 12 drives, 4 of which can be swapped out for a tape system. These systems use high-speed 2.5-inch drives spinning either at 10,000 or 15,000 rpm. On the Lenovo, these drives had a capacity of 300GB, so configuring these servers with multiterabyte high-speed storage is entirely possible. This means that sites can add server capacity without necessarily increasing the network traffic or the number of network storage devices.
We examined the Dell PowerEdge R710, the Hewlett-Packard DL380 G6, and the Lenovo ThinkServer RD220. These Dell and HP models are both well-established IT workhorses, while the Lenovo is mechanically and electrically identical to the IBM System x3650 M2. All three models are 2U-sized rack-mounted servers. As with pizza boxes, you access their contents by opening and removing the lid. These servers, like most models available today, are superb at making component swaps and other service tasks easy. All three vendors enclosed a map of the server components with various key points of information that they affix to the inside of the cover panel.
Inside the chassis, all adjustments and parts replacement can be done using essentially no tools. Five years ago, the tool-less chassis design was a big and welcome innovation. Today, it is standard fare. The greatest challenge is generally removing the cowlings that direct the fans' air flow when it's time to add RAM or insert PCI riser boards. In all three of these servers, accessing the PCI slots proved fairly straightforward, although the HP requires removing a “riser cage” into which the boards were fitted. The entire cage is then re-inserted into the corresponding slots. This was the only nonintuitive physical manipulation of the servers I encountered.
The greatest variation in these systems is how they notify admins of defective components. Only the Dell PowerEdge R710 has a front-panel LCD (limited to one line measuring roughly 1.25 inches in length) that shows error codes. It is surrounded by small buttons that let you traverse a tiny menu of admin items on the LCD panel. The Lenovo makes original use of system-warning lights. If one of these lights should flash, the admin slides out a 3.5-by-1.5-inch panel that contains a two-digit display and a series of tiny LEDs. The LED corresponding to the specific component lights up when a problem occurs, and the error code appears in the two-digit display.
We found this design awkward to use and wondered whether a system at the top of a rack could be diagnosed without forcing an admin to get a ladder to read the slide-out panel. Lenovo points out that this approach provides great specificity to service personnel who can immediately swap out the defective part. For that purpose, we prefer HP's approach, which has the same component-warning LEDs directly on the front bezel. There is a light for every processor, DIMM, disk, power supply, and fan. One scan of a whole rack of servers can tell you if any of them has a defective component. For purposes of such quick scans, HP's approach is easier than even the Dell LCD panel.
Oddly, the Lenovo's start/stop button is embedded in the side of the diagnostic panel. This gives it a distinctively mushy feel, which wouldn't be worth a comment except for the Lenovo's very long delay in starting up. On the Dell and HP, you can start the system seconds after you plug the server into the power outlet. On the Lenovo, the wait is much longer; it was hard to time exactly, but because it went far past the industry standard, the temptation was to think the machine was DOA. Lenovo knows about the frustration this delay causes and will fix it in a future release.
A final point before we look at benchmarks is that the Dell server, surprisingly, did not come with redundant power supplies. On both the Lenovo and HP models, dual power supplies is the standard delivery model. On the Dell, the second power supply has to be specifically ordered — it's not the default. IT sites considering the R710 should add the cost of this power supply when comparing prices.
The test results
The comparison table shows three performance benchmarks. Taking these in reverse order: The VMmark benchmark determines how good the system is at supporting virtualization. This is a complex benchmark (downloadable from VMware at no cost) that requires the use of multiple systems and the expertise of performance engineers. Because of this, InfoWorld relies on the vendors' test results, which are examined and posted by VMware. These results were obtained on machines with slightly different configurations than those I looked at — namely, they all had 96GB of RAM. Lenovo does not run the VMmark benchmark, so we used the result for the IBM System x3560 M2, which, as mentioned previously, is the Lenovo's mechanical and electrical clone. The VMmark results show essential parity between the three models (less than 2 percent variation from top to bottom scores). This is a particularly good result for the Lenovo machine, as it uses a slower processor.
The SPECjbb 2005 benchmark (which can be purchased from SPEC.org) builds data warehouses in RAM and does a series of analysis on the records. It primarily measures the performance of Java business operations and memory access speeds. In this test, the Dell and HP were essentially tied, while the Lenovo lagged, primarily due to its slower but less expensive processor.
Finally, we ran the freely downloadable Stream benchmark, which was recently added to InfoWorld's complement of server tests. It measures memory bandwidth in a parallel processing context — a useful test in light of the many cores these systems now support. On this benchmark, the HP holds a slight performance lead over the Dell (as it did with SPECjbb_2005), while both systems led the Lenovo by roughly a 20 percent margin.
As to power consumption, the Dell R710 sipped power when running at 0 percent load (but not hibernating), roughly 63 percent of the HP DL380's consumption — a substantial difference, especially in the use case of servers that need to be on all the time but see only infrequent activity. The Lenovo ThinkServer, whose processor consumes less power overall, fell between the HP and Dell at 0 percent utilization. At 100 percent load, the Lenovo ThinkServer was the stingiest on power consumption due to its processor model; between the HP and Dell, which use the same higher-end CPU, Dell again was the power-savings leader. (Note that power consumption of each system was measured using only one power supply.)
For sites expecting to use the internal hard disks, it's worth noting that HP's RAID controller, the HP Smart Array P410i, runs at 6GBps for SAS drives, while Dell's PowerEdge RAID Controller (PERC) model 6i works at 3GBps. In both cases, the RAID controllers can be upgraded or downgraded. The Lenovo's RAID controller was harder to establish due to the use of both Lenovo and IBM part numbers, but it appears to be 3GBps, upgradable to 6GBps.
The report card
The report card for the three systems shows patterns that in some ways reflect the features of the scoring system, rather than the devices themselves. When viewing the scores, note that InfoWorld publishes the weightings for each component in the overall score so that you can judge the servers on the parameters as weighted within your organization. One parameter that requires further explanation is the category of value, which represents our view of the price. We use the retail price for the system configuration sent to us. On this review, however, HP and Lenovo both used Microsoft Server 2008 Standard Edition, while Dell sent the Enterprise Edition. The cost difference in editions is significant. In the case of Dell, it increased the price by $2,200. For a relevant price comparison, we repriced the Dell server for the Standard Edition, which is what you find in the accompanying table.
Essentially the Dell and HP systems are tied for first place. The disparity in scores is due to only one factor: power consumption. All other scores are identical. Given the similarity in both the feature sets and the benchmark results, the scores should look a lot alike. The HP has slightly better performance and less favorable power consumption and pricing. How you value these difference will decide which model is for you. It's clear, however, that either way you'll be well served.
The Lenovo ThinkServer slightly lags its two counterparts, but this is due primarily to the difference in performance attributable to the company's use of the E5540 models of Intel's Xeon processor. Had our ThinkServer been outfitted with the same X5570 chip as Dell and HP and scored comparable performance results, it would be tied with the HP ProLiant. Lenovo's choice of processor enables it to be the value leader in this category. There is no doubt that organizations that want to expand computing power without gutting their capital budget will find the Lenovo to be their best option — and so will small businesses.