IBM has been a top performer in recent server tests, which is why, for this hands-on exploration of blades vs. rackable servers, we tested IBM's state of the art in both categories - the HS21 and HS21 XM blades and the x3550 and x3650 rackables - and focused on performance, power consumption and manageability.
Because blade offerings have at least some things in common, we believe these test results are applicable generally to blades from other vendors - Dell and HP, for example - though there may be vendor-specific considerations in assessing products from others.
We found that blade servers reduce but don't eliminate redundant hardware components requiring electrical power, a pivotal consideration in total cost of ownership (TCO). That said, while IBM's blade servers equalled the performance of its rackables, the blades were more power efficient and potentially much easier to service.
While rackables are less expensive than blades when buying only one or two servers, because you have to factor in the cost of the blade chassis, a fully configured blade chassis is the more economical hardware buy than purchasing the same number of rackable servers. On the other hand, adopting a blade configuration demands a vendor lock-in that rackables don't, because they can be added one at a time as needed. Also, blades fall short for applications that require large amounts of on-board storage.
We tested two types of blades from IBM, the HS21 and HS21 XM. Both slid into the IBM Blade centre H (BC-H) chassis comprising a 10GB Ethernet switch, a 1GB Ethernet switch and four power supplies. The XM had a 2.33GHz Intel quad-core CPU, and the HS21 had a 2.0GHz CPU. IBM sent the XMs with 16GB of double-data-rate, second-generation (DDR2) memory, and the HS21s with 8GB of DDR2 memory.
In their specifications, these blades are comparable to the two x3550 rack servers that IBM sent, except that there's more space available inside the x3550s. IBM also sent one x3650 2U server, which included three RAID drives in RAID 5 configuration, a RAID controller and a hot-spare drive.
Prices for the blades ranged from around £3,500 for an HS21 with 2.0GHz twin-quad-core Intel CPUs with 8GB of memory, to around £5,000 for a hefty HS21 XM with a twin set of Intel 2.33GHz quad-core CPUs, 16GB of memory and an onboard RAID 0/1 configuration. Additionally, the BC-H chassis as specified costs about £9,000.
The rackables cost £4,000 to £5,000 depending on CPU clock-speed enhancements, extra memory and RAID controllers with extra drives.
Blade servers' drawbacks still prevent their adoption for some applications, such as those requiring large amounts of on-board, indeterminate storage expansion.
Even though performance tests among servers with comparable CPU clock and memory configurations were virtually identical, faster CPUs probably will arrive first in rackable form. This may not always be true, but our experience suggests any vendor can adapt the 1U or 2U form factor rapidly, while blade vendors have to plan in advance to accommodate new motherboard component combinations.
Blades' density sometimes works against them. They weigh more than rack-mounted servers. The backplanes of blade servers also represent potential (but rare) single points of failure - despite redundant power supplies and other duplicated or redundant components.
In addition, a datacentre using blades is captive to the vendors' business partners to supply such devices as storage-area network (SAN) switches or Ethernet switches, because these must be housed inside the blade chassis. This is OK if the vendor is competent, spares are available, and the chosen technology components mesh with current and proposed network operations centre (NOC) gear. If all works with little orchestration, costs will remain low.
Performance is a wash
Performance among the IBM servers we tested was startlingly similar. All had Intel quad-core CPUs - two per server - and the performance variables we saw favoured RAID 1 and RAID 5 configurations in terms of speed.
Our usual benchmark, LMBench3, showed that one of the tougher tests -- processor fork+execve (which deals with file I/O and memory shifting) -- executed in an average of 173.2 microsec vs. 289.9 microsec, a 40 percent increase in speed.
The results of all tests on all IBM servers were almost identical -- except that the RAID drives on the HS21 XM models and x3650 server were slightly faster than NT-File-System- or ext3-formatted drives. The faster, 2.66GHz CPU clock of the x3550 and x3650 units gave them a slight performance edge as well.
Comparing the performance characteristics of these servers, therefore, is an exercise in apples-to-apples, if you'll pardon the testing pun.
It all boils down to power
Often, not enough attention is paid to what may be the largest server operational cost: power consumption.
IBM's stated power budget for the blades and the rack servers was higher than we measured. In our test, the low ambient temperature (our lab operates at 68 degrees Fahrenheit) and the fact that IBM could not supply us a fully loaded blade chassis may have contributed to lower-than-expected power consumption. We multiplied our measurements to obtain a comparison between 14 blade servers in the IBM chassis and 14 IBM rack servers.
If datacentre cooling and power and the centre's floor load-bearing qualities (the floor has to bear loads that literally can weigh a ton) are satisfactory, a blade chassis filled with blades provides the least expensive TCO in terms of power consumption over five years (see power-consumption graphic).
Rack servers, compared with the components in the blade server and a fully loaded blade server, consume more power, and this adds to their infrastructure expenditures over a five-year life. The flexibility of having RAID 5 in a rack server is offset by its power consumption, although with a blade configuration you might run into additional costs for external storage.
The cost of the IBM blade server configured with the maximum 14 blades is US$117,400. Add power consumption (except for the cost of cooling), and the total becomes $123,900. The equivalent number of x3550 1U servers at $7,900 comes to $116,700. When the cost of the needed six-port 10G Ethernet and 24-port Gigabit Ethernet switches and power consumption costs are added, the cost climbs to $129,200. The x3650 has a total capital cost of $129,800 ($9,300 each), and with the aforementioned 10G Ethernet and Gigabit Ethernet switches plus power consumption, the cost comes to a whopping $150,700.
In the widest stretch, between the loaded BC-H vs. an equivalent number of x3650 servers, there is an 18 percent price differential over five years. This becomes amplified if power costs increase over our - admittedly very inexpensive - three cents per kilowatt-hour, which is a typical price for a large NOC in the Midwest, where we tested this gear. In the UK, powerIn other parts of the country, power can be more expensive and subject to peak use rates.
Servicing blades ultimately is faster than servicing rack servers for several reasons. First, the overall size and weight of a blade is easier to handle, because its power supply is common to other components. We could pull a blade from a server in about three seconds, whereas pulling a rack server took more than a minute, even when cable-management components were well designed and implemented. It would take much longer if a rack were in any state of disarray.
We measured how long, once a unit was out of a rack, it took to change memory or a hard drive - the two components most frequently changed by a server administrator. In this case, the times were much the same. The time needed to reinsert equipment and bring a server back online was far faster with a blade server than with a rack server, because no cables needed to be reconnected.
There is much to be said for how using blades reduces the number of power cables needed, as power cables are reduced from as many as 80 to eight. The number of copper or fibre Ethernet cables also is reduced dramatically, as well as the clutter that cables introduce and the air-flow reduction that they represent.
Additional costs are associated with management components. In the case of the IBM servers we tested, we used IBM Director to manage both types. Director must be "housed" on its own rackable server or blade, and the server has its own set of costs, though it need not be a high performer with lots of storage.
We downloaded various operating systems, including several instances running via EMC's VMware ESX, and configured them on the blade and rackable servers using Director. Some vendors charge per installation or per server for management components, and it's important to check how much these valuable and time-saving applications will cost.
The amount of onboard blade-server storage media is limited, a function of the small size of a blade and its heat dispersion needs. The blades we tested could have RAID configurations on them, but only a RAID 1 mirroring or RAID 0 non-redundant disk striping. This means that capacities are bound by the size of the drives vendors ship with the blades, in this case, a limit of 288GB (maximum; we tested two 73GB drives in the blades) from IBM. Some will argue that 288GB is a lot, but it's a ceiling amount unless iSCSI, Fibre Channel or other SAN connection methods are onboard and well configured.
By contrast, the 2U x3650 can house a RAID 5 configuration internally by using four drives - three for RAID and one as a hot-spare. These drives, like those in blade-server housings, can be changed to accommodate larger storage needs, but it's unlikely that administrators will change drives once a unit is in service for a production application.
With more native space available, rack-mounted servers have a higher denominator of native storage capability. Some vendors are putting this same storage capacity in 1U form-factor rackable servers as well. There's a tendency not to change components in a working system even if better components will achieve a longer application or service profile life.
The result is that blades need to connect to an external SAN if it's known or perceived that future storage-growth needs will exceed the amount of space available on the blade. For organisations that have SANs, blade-server attachments are well understood, and a variety of Fibre Channel SAN gear is available. If there's no SAN nearby, however, iSCSI can be used, subject to the bandwidth consumed over Ethernet connections by the virtualised iSCSI-based traffic.
It's also possible to put additional I/O cards into blades, as well as into rackables. The IBM line uses PCI and PCI Express (PCIe) cards in its rackables. Options to increase to two cards makes rackables more flexible in terms of I/O card expansion. The memory options in blades and rackables are similar.
Another potential disadvantage to blades is commitment. If you choose a blade-server vendor, your organisation becomes captive to that vendor's service policies, component availability and service organisation. This isn't necessarily the case for 1U deployments, because the 1U space can be occupied by any vendor, but a blade space will be occupied by a server from the chassis' manufacturer. Relationships become tighter if a blade environment is chosen.
Blade density represents a huge pool of computing power per cubic inch. Typically, four 9U IBM blade chassis will fit in a 42U rack with space to spare, which means almost 20,000 watts of power to cool. Forty x3550 rackable servers in the same rack use at most 75 percent of that wattage. There are 56 blades in this configuration, however, vs. 40 rackables inside a single rack, maximised in this way.
Performance was essentially the same between rackables and blades. Blades are easier to manage and service. We believe there's a vendor-captivity element to purchasing or deploying blade servers, as well as a decided limitation on local storage. However, in an organisation with very good datacentre infrastructure, if you can live with the vendors as in-laws, you'll love the blades.
Henderson is principal researcher, and Dvorak, a researcher at ExtremeLabs in Indianapolis, USA.