Buyer's Guide

Commodity Superservers: Muscle-bound PCs

by Jay Milne

PCs on steroids--this phrase attempts to describe the latest contender in the world of client/server computing: commodity superservers. While taking advantage of new PC technology, commodity superservers are starting to displace the traditional midrange systems used for mission-critical applications. One major

midrange systems vendor estimates that commodity superservers are replacing up to 20 percent of their new purchases. (See chart at right on page 124.) What applications are being run? How about SAP, Lotus Notes or SYBASE SQL Server on NT, or 1,000 users on NetWare 4.1? The decision to purchase one of these systems is driven by the need for a specific application. But what should you consider when selecting these systems? There are three interrelated core components: performance, fault tolerance/availability and growth.

Performance Performance was the biggest advantage that midrange and high-end servers had over traditional PC systems. Combining a single 486 processor, an ISA bus and no operating system that would scale tended to constrain the PC architecture for use solely on the desktop or in a low-performance file/print services function. The arrival of the Intel Pentium, high-performance I/O buses, such as PCI, along with operating systems and applications that take advantage of multiprocessors, have made the commodity superserver a viable platform for mission-critical environments.

To achieve the level of performance required of these systems, mu ltiple CPUs are necessary. To fully use the power of multiple CPUs, both the operating systems and the application must be designed for symmetrical multiprocessing (SMP). An application that's not designed for multiprocessor (MP) systems won't run noticeably faster on an MP server than on a single processor system.

Multiprocessing can take several forms. The most commonly implemented architecture for low-end client/server systems is SMP using a dedicated cache for each processor. The primary difference between asymmetrical multiprocessing (AMP) and SMP, is that with an AMP system, each processor is dedicated to a certain task or function, such as I/O or running applications. With SMP, every processor in the system is available to run any process, allowing for more efficient use of the processors. A high-performance commodity superserver is very different from those systems that are dual CPU systems intended for the desktop. While having dual Pentium 133-MHz CPUs will increase performance, they tend to use a shared cache design and aren't fully optimized to take advantage of all the computing power available.

Performance increases are not linear with respect to the number of processors added. A four-CPU system is not twice as fast as a dual CPU system, due to other system constraints. This is where many hardware vendors can value-add. A commodity superserver needs a high-bandwidth (usually 64-bit or 128-bit wide) system bus.

One of the most significant factors to affect MP systems is cache. Cache is high-speed and high-cost memory that stores the most recently used data. Most MP systems use a two-level cache design, in which a small amount of cache resides on the processor itself and a large L2 cache is connected directly to the processor. Some systems, such as Compaq Proliant 4000, utilize an optional L3 cache to help increase performance. The effectiveness of an L3 cache is in heavy debate and many industry experts agree that a well-designed L2 cache can support an MP system. How the cache is designed is also important. Many dual-processor PC systems use a shared cache design, where both processors share the same L2 cache. This is the low-cost, lower-performance method. Higher-performance systems use a dedicated cache architecture, where each processor has its own cache. The P6 from Intel is incorporating a new design where either a 256-KB or 512-KB L2 cache is on the processor itself, thus eliminating the need for system vendors to engineer an L2 cache. This feature, along with many others, makes the P6 significantly faster than current Pentium systems.

For I/O, PCI is becoming the de facto standard, bypassing EISA as the high-performance bus of choice. While there are more EISA adapters on the market, allowing for greater choice, PCI is better suited for the multiprocessor environment and is capable of 132 MB per second as opposed to EISA's 32 MB per second burst rate. The ISA, a legacy of the old PC/XT is not at all suitable for a server environment and should be avoided. To increase the number of adapter slots, many vendors are implementing dual PCI buses either via a PCI bridge or by connecting directly to the system bus in a peer fashion. The latter is considered the better alternative.

Some system manufactures, such as Tricord and NetFrame, use a proprietary disk I/O bus that allows for very high performance. System vendors are using intelligent subsystems that have dedicated CPUs that do nothing more than process I/O functions. Systems such as NetFrame utilize Intel 486 processors to increase performance, and Compaq uses a 486-class processor on its SMART Array RAID controller. By utilizing a separate processor for I/O, the main CPUs aren't burdened with those tasks.

When considering disk storage and adding optical and tape drives, SCSI is the only viable option. Avoid IDE/EIDE, as it does not offer the multitasking, support for a wide variety of devices nor the performance of SCSI. Current versions of SCSI can achieve up to 40 MB per second in burst mode. In the near future, alternatives to SCSI will appear, but for now, SCSI is the most versatile and widely available solution.

Fault Tolerance/High Availability The goal of any system is to remain operational for as long as possible, but a plethora of problems can cause a server to crash, ranging from software faults to failed hardware to operator error. According to a study by Intel, hard disk and power supply failures are the most likely components to fail. (See the figure at left.)

It's no surprise, then, that all the high-end systems utilize some form of disk fault tolerance. The most common implementations and the minimum that a system should support are RAID levels 0, 1 and 5. While RAID level 0, often called disk striping, has no fault tolerance, most systems allow for the RAID 0 disk array to be mirrored (also called RAID level 1). This configuration offers the highest performance because data can be written and read to both disk channels and to multiple disks at the same time. RAID 5 is also popular and stripes both data and a parity value, which is created by performing an XOR function on the data across all the drives in the array. A minimum of three drives is needed for a RAID 5 array. Additionally, the RAID subsystem should support an on-line spare, which allows for a spare drive to replace a failed drive automatically and hot swapping capability, which allows for the removal of any drive without having to shut down the system. Disk fault tolerance is essential in any server environment besides backups. Some people mistakenly eliminate total system backups, thinking that a fault tolerant disk subsystem is enough protection. A RAID system doesn't reduce the chance of total system failure due to a natural disaster, accidental file deletions or theft. Also consider storing the backups either in a different location in your facility or, better yet, offsite.

Systems employ other features, such as redundant CPUs. NetFrame uses a technology called Gemini mode, which enables two CPUs to execute instructions in lockstep mode. If one processor detects an error with the other CPU, that CPU is disabled. More common features are redundant power supplies (make sure they load balance between them) and redundant fans.

Other facilities that help maintain maximum availability are remote server monitoring and management. This allows the server to send alerts to a central management console.

Consider technical support, servicing and the availability of replacement parts. What kind of technical support do they have? Is it 7x24? Do they have priority support that allows you to bypass the front-line support personnel? What are the hidden costs? Inquire about servicing of the unit. If your corporation is multinational, ask about worldwide support. If the server you choose only uses proprietary parts, get a guarantee about part availability.

Growth Finally, a key superserver element is its ability to handle future demand. The server might currently be supporting only 200 users, but next year you may need support for 500 or 1,000 users. Four components should be covered to ensure growth potential.

First, you'll need memory. The superserver should be capable of at least 512 MB of RAM, and a good range is between 512 MB and 1 GB, since applications like database servers run better with more RAM. Also, as disk size grows, the need to cache the data and file system grows as well. A common rule of thumb for a NetWare server is 16 MB of RAM for each 1 GB of storage. If you have a 40-GB disk array, that works out to be 640 MB of RAM! You should choose RAM that has some error-correcting capabilities, because even though it costs more than ordinary RAM, it allows for detection and correction of multiple bit errors.

The I/O bus is the second key area. The most important question is, "How many expansion slots are there?" Avoid any system with less than three available slots and look for systems with five or more. This allows for the addition of network cards and SCSI cards for extra disks and optical devices that could b e used for data migration or data storage.

The third area is the number of CPUs supported. Look for at least a four-way SMP system. For NT and NetWare SMP, anything above 16 processors isn't worth it. According to Intel, the sweet spot for commodity superservers is around four to six processors. Some systems allow for more, but are typically more expensive due to the additional engineering required to support the additional processors. Also, make sure you can replace the CPUs with faster ones.

The fourth component is disk storage. How many GBs can the system support? If you need more GBs, how do you do it? While there is no product on the market that allows for the dynamic expansion of a RAID 5 volume, products are coming down the pike, such as the new HP AutoRAID. A disk array should support 4-GB drives and should support at least 14 drives.

Now What? The implementation and deployment of a client/server system are not merely a matter of purchasing hardware, installing the software and letting users go at it. Performance, fault tolerance/availability and growth are all key elements, but other supporting issues must also be taken into account. The total cost of a client/server solution is more than the cost of the hardware and the software.

Jay Milne can be reached at jmilne@nwc.com.

October 15, 1995