Realizing the Need for Speed
May 17, 1999
By Dave Fetters
The enterprise server market has stepped up to meet the high-performance demands of server applications. Complex Web servers, software encryption, virus scanning on the fly, and video and voice compression are just a few applications that let a true high-performance server flex its muscles. Faced with these high demands, server vendors are rising to the occasion with blazing fast servers that offer increased reliability and scalability.
This past year, we saw an increasing need for denser, more serviceable servers. Fortunately, vendors answered the call; most now provide both rack-mount and tower versions of their servers. Even workgroup servers have been redesigned for rack-mounting. On the performance side, it's grown increasingly difficult to pick a clear winner based on our comparative tests of Intel-based servers. In many cases, the margins of victory are nearly imperceptible, with vendors now using very similar technologies and hardware. Features lists for servers are also beginning to blur and it's getting increasingly difficult to tell products apart. Makers of RISC servers have also taken heed of clustering demands by providing higher-density servers. And RISC-based server prices are becoming more competitive, which puts a tighter squeeze on the midrange server market.
Not surprising, reliability and fault tolerance are getting more attention. The hot new marketing phrase is "five 9's reliability" referring to 99.999 percent uptime. Server makers that promise such 24x7 operation are becoming more appealing as the cost of downtime rises. In reality, however, servers still fall short of the 99.999 percent mark. Hot-swappable components are an integral part of the 24x7 formula and they're starting to get a lot of attention, largely thanks to the hot-swappable PCI bus.
In addition to PCI hot-swappability, many servers now incorporate 64-bit PCI slots. With 64-bit PCI, peripherals can access up to 4 GB of memory at a transfer rate of 266 Mbps. Gigabit NICs, and newer dual-channel RAID controllers and SCSI cards support 64-bit PCI.
CPUs Gain Strength
There is big news coming in the area of server-class CPUs. In June, AMD will release the K7, its first server-based CPU, which boasts several key improvements. Although AMD's chips have never had a reputation for superior FPU (floating point unit) performance, the vendor is stepping up to the plate this time. The K7 will incorporate three out-of-order, fully parallel FPU pipelines. Additionally, the K7 will use Digital Equipment's Alpha "EV6" bus protocol, which was originally designed to scale up to 256 processors. To further support the multiprocessor environment, the K7 will boast up to 8 MB of L2 cache. The added L2 cache decreases the CPU's need to saturate the bus and should enhance performance scaling from AMD's line, a notorious sore spot. This large cache size also will allow for a cacheable memory limit of up to 65 GB. To support high clock speeds, the K7 will feature 128 KB of L1 cache. Although the K7 could turn out to be AMD's crown jewel, the company is infamous for bringing products to market late, and sometimes not at all.
Intel is also in the midst of the fray with its latest Pentium III processor. The PIII sports an all-new multimedia instruction set called SSE (streaming SIMD extensions) and a higher clock speed than its predecessor, the Pentium II. But it remains to be seen if SSE-optimized applications will justify the SSE upgrade. For those buying servers, the PIII may be a non-issue, since SSE does nothing for server performance. But it does offer higher clock speeds and does a better job at scaling in a multiprocessor environment. Our first look at the Compaq 6400R proved that a four-way Pentium III is no joke (see "Compaq 6400R: Pint-Sized Powerhouse," www. networkcomputing.com/1010/1010sp2.html). In June, Intel will further prove its scalability by shipping its eight-processor Profusion chip set. Intel recently showcased another CPU, code-named "Foster," that is basically a Pentium III running at 1,002 MHz--overclocking at its finest. Another new CPU, code-named "Coppermine," is expected to debut in the latter half of the year. It will feature a shrunken .18 micron die and a 133-MHz front-side bus.
But the Intel chip that's creating the most buzz is Merced, the company's first 64-bit processor, which is slated for release sometime in the first quarter of next year. There's an interesting twist to Merced: Intel has entered into an agreement with Sun Microsystems to develop a version of Merced optimized for Solaris. Not surprisingly, speculation about just when Merced will ship is rampant. Unfortunately, much of the undercurrent of speculation surrounding Merced focuses on further delays and subpar performance. Intel officials remain optimistic in public, and despite unfavorable rumors, say the chip's on track for release.
RISC technology has continued to impress us, with the Alpha 21264 running at 600 MHz and the soon-to-be-released Apple G4 and Sun's UltraSparc-III debuting at 600 MHz. Increased bus speeds incorporating advanced-switched bus fabrics continue to demonstrate RISC servers' superiority in speed and scalability. Alpha's upcoming EV6 bus architecture is designed to support up to a 400 MHz bus. We'll round up a handful of RISC servers and gear them up to do battle in head-to-head tests later this year.
Thin Clients Ride Again
With all this power, there's hope again for thin-client technologies. Over the past year, system management and total cost of ownership have drawn enough attention to warrant a closer look at thin clients. The thin-client area is thick with Windows CE-based terminals, not to be confused with CE-based handheld devices. When we tested a handful of WBTs (Windows-Based Terminals) late last year, we found performance to be merely adequate, while the inherent shortcomings of Windows CE make it difficult to lock down configurations (see "Windows-Based Terminals: Construction in Progress," www.networkcomputing.com/921/921r1.html). In addition, they lack power management and wireless capability.
More positive is that with Windows Terminal Server and a host of mature ICA (Independent Computing Architecture) clients, the pieces are starting to fall into place. One major drawback to the widespread adoption of WBTs has been price. With PC prices plunging, many IS purchasers have chosen to play it safe by buying PCs and then hooking them up to the usual fat server-side applications.
These applications will need beefy storage, of course. Thankfully, storage prices have dropped and performance will get a boost from SCSI Ultra3, which increases throughput from 80 MB to 160 MB. Servers will also see a hike in out-of-box storage, or more specifically NAS (Network-Attached Storage) and SAN (Storage-Area Networks). NAS servers are beginning to receive the attention they deserve. For the same reason that routers evolved from software on a server to a hardware box, I/O-intensive file operations are being transferred to a NAS server to cut overhead and free up resources.
SANs are making big headlines everywhere. A SAN is a specialized high-speed network that interconnects different kinds of data storage devices with associated data servers. The interconnect between the servers and storage are formed by the 100-MB-per-second fiber-channel technology and are typically built around a fiber port hub. Still in its infancy, SAN technology is undergoing heavy development. SANs only recently made the migration from an arbitrated loop to switched-fabric hardware.
SANs offer several key benefits, starting with their role in disaster recovery. A SAN can span 15 kilometers, making off-premises storage feasible. Many times the data storage devices are RAID boxes. Since the SAN is basically on its own network, these RAID boxes can easily replicate data between themselves without ever having to go over a LAN. Additionally, each server within the SAN can realize a huge performance benefit from all of the RAID boxes working in conjunction. If that's not good enough, the RAID sets are not physically bound to a storage device. With a switched network, it's possible to create a RAID set that spans over a WAN. A SAN also creates a heterogeneous data store using a transparent metadata scheme. Since a SAN can be rather complex, it's not surprising that management software is extremely important.
|
Here
Here
REPORTS
ANALYTICS
Follow 
