Our tests of Opteron and Itanium servers show big performance gains today and promises of bigger gains to come
With the arrival of the AMD Opteron and Intel Itanium, commodity servers built on these processors have joined proprietary RISC systems from IBM, Sun Microsystems, Hewlett-Packard, and others in the 64-bit landscape. With prices starting at just over $2,000, Opteron and Itanium systems — running Linux or Windows — are already carving out a niche in high-performance computing clusters, where they are used to run compute-intensive scientific- and financial-modeling applications. Eventually they will replace their 32-bit forebears in corporate datacenters, and clusters of them may even challenge 64-bit Unix systems costing hundreds of thousands of dollars.
How long this will take depends on software vendors, who must rewrite their applications for the new 64-bit CPUs. Many operating systems are already available for Opteron and Itanium. In addition to Red Hat Enterprise Linux AS 2.1, which supports Itanium, and SuSE Linux Enterprise Server, which supports both Itanium and Opteron, there is an Itanium version of Windows Server 2003, and Itanium and Opteron versions of Turbolinux Enterprise Server 8. Databases such as IBM DB2 and Oracle 9i, and application servers from IBM, Oracle, and BEA Systems should begin to be available this year. But enterprise app vendors, with the exception of SAP, have been slow to commit to a road map of support.
Ultimately, the appeal to business software vendors and customers alike will depend on performance. Compared to the previous generation of 32-bit systems based on AMD’s Athlon MP and Intel’s Xeon MP, the Opteron and Itanium CPUs support much more memory per processor, higher-speed connections between memory and the CPU, and faster interprocessor communications. To what extent do these advantages translate into real performance gains? To find out, I looked at Web server performance on four 64-bit Linux systems, including an Itanium 2 server from HP and Opteron systems from Appro, Newisys (a company partly funded by AMD), and Pogo Linux.
There are many other types of benchmarks that could have been used here, but all are subject to various problems, including finding code that will run on both Itanium and Opteron processors; that is optimized for 64-bit operation; and that is not associated with some major player. Because system and database benchmarks are typically oriented to a specific type of hardware or operating system, I compared the systems based on Web server performance, using the industry-standard Apache Web server, which was included as part of the standard release of Linux on all four systems, as well as Zeus Technologies’ Zeus Web Server.
Load testing these Web servers on the four systems produced interesting results. First, Web-server performance on these systems was substantially better than on the 32-bit Xeon hardware I compared them against, suggesting that IT shops would benefit from migrating their Web server farms to 64-bit systems today. Second, Apache performance on Itanium was abysmal compared to Zeus performance; if you’re considering Itanium for Web serving, go with the Zeus Web server. Third, the performance differences between Apache and Zeus on Itanium point to the need for software that not only runs on 64-bit hardware, but is optimized for it.
Finally, regarding the four systems tested, all proved to be solid machines. Among the Opteron units — all 1U boxes — the Appro and Pogo Linux systems offer great value and horsepower for the price, while the Newisys offers a more substantial feature set, including a management card and hot-swap SCSI drives. The Itanium-based HP system, a 2U box, offers great expandability and redundant everything, albeit at a higher price.
Three Little Opterons
The Opteron has one advantage over the Itanium: It can run 32-bit applications in native mode, while the Itanium runs 32-bit applications in emulation mode. This means legacy 32-bit applications will run slower on the Itanium than on Xeon systems or Opteron. It also means that applications can be ported to Opteron gradually, weaving in 64-bit support component by component, without requiring a wholesale rewrite before deployment.
Each of the three Opterons I tested ran SuSE Linux’s Enterprise Server, while the Itanium system from HP ran Red Hat’s Enterprise Linux AS. All four systems came with Apache installed; I also installed Zeus Web Server 4.2r2, which is optimized for both Opteron and Itanium.
The Appro 1122H is a value-oriented 1U server, available with two 1.6GHz or 1.8GHz Opteron processors, up to 16GB of RAM, one or two ATA or SCSI drives, and two 10/100/1000 Ethernet interfaces. My test unit came with two 1.6GHz processors, 2GB of RAM, one 80GB ATA hard drive, and SuSE Linux Enterprise Server already installed.
The Appro system produced the best performance by a slight margin across all my tests — despite having the lowest price tag of the bunch. The case is nicely put together and required no tools to gain entrance. It was the only system that included rack-mount rails, which were well-engineered. On the downside, the power button is tiny and hard to press, and the keyboard and mouse ports were not labeled.
The next-best performer was Pogo Linux’s PerformanceWare 1264. Pogo Linux has been specializing in Linux systems for several years, and the PerformanceWare 1264 is typical of their product line — well-engineered with no frills. At $2,489, it’s nearly as affordable as the Appro system, and it’s solidly in the middle of the Opterons in performance. As tested, the system included two Opteron 1.6GHz processors, just 1GB of RAM, a 40GB ATA hard disk, and two 10/100/1000 Ethernet ports.
The PerformanceWare was the only one of the four systems that required a screwdriver to open the case, and the only one with drives that couldn’t be removed without opening the box. The interior of the system is laid out well, with solid airflow controllers and dual fans.
Interestingly, although the Newisys 2100 Server had faster processors, a faster hard drive, and more RAM than the others, it recorded the lowest performance of the three Opteron systems, though only by 2 percent to 3 percent in most of the tests.
Partly funded by AMD, Newisys has been working in close cooperation with AMD to create reference motherboard and chassis designs. Newisys systems are not available directly from the manufacturer but can be purchased from a number of resellers, including Colfax International, ProMicro, and RackSaver.
The Newisys 2100 has a surprising number of enterprise-class features for a system under $3,000. These include hot-swap SCSI drives, four Ethernet interfaces, an extra PCI-X slot, and a management processor.
Accessible through its own Ethernet interface, and running a separate Linux kernel, the management blade can reboot the system, monitor system components, and inventory hardware and software on the system. A small LCD on the front of the Newisys box shows the status of the management system and allows control of its basic features. If the management processor is set to get an IP address from DHCP, the LCD shows what the address is, making the process of getting to the system the first time much easier.
Pulling up the rear in my Web server performance tests, but not far behind, was the Itanium-based HP Integrity Server rx2600, a 2U, enterprise-oriented server with lots of expandability and redundancy. My rx2600 came configured with dual 1GHz processors, 4GB of RAM, three hot-swap 36GB Ultra320 10K SCSI drives, a RAID controller, four Gigabit Ethernet ports, two 10/100 Ethernet ports, HP’s iLO (Integrated Lights-Out) management processor, dual hot-swap power supplies, hot-swappable fans, and Red Hat Enterprise Linux AS. This system will soon be shipping with the newest 1.3GHz and 1.5GHz flavors of Itanium; it’s also available with HP-UX or Windows Datacenter Server.
The management processor allows full access to the system, including the ability to perform BIOS upgrades, monitor all internal components, power the system on or off, reboot it, and inventory hardware and software. One slight negative: The rx2600 uses USB ports for keyboard and mouse, which could cause problems for datacenters with existing KVM-switch infrastructures.
To gauge the performance of these systems, I tested the loads the Apache and Zeus Web servers would support before bogging down, using RadView Software’s WebLoad 5 running on Ixia’s TXS4 Load Module to simulate up to 1,000 simultaneous users. I also ran the same Web servers on a dual Xeon 2.8GHz server to create a baseline. The table below shows the number of virtual clients that were required to generate round-trip times of over than five seconds with Zeus and over two seconds with Apache.
These results are a great advertisement for the Zeus Web server. Not only did Zeus respond much faster than Apache for a given number of clients, it handled much greater loads than the Apache server — even 1,000 virtual clients only produced four-second response times. In addition, it’s stable at much higher loads than Apache, and even at high loads produces very uniform results: Raise the number of virtual clients and the response time goes up in a very predictable fashion. With Apache, on the other hand, response times, hits per second, and throughput numbers varied wildly with loads starting at 100 clients. Zeus produced very uniform results up to 1,000 clients.
With the Apache Web server, the Opterons showed a clear advantage. The HP Itanium system started producing long round-trip times at a load of fewer than 300 clients, while the Opterons were able to sustain loads above 800 with reasonable round-trip times. When running Apache, the Pogo Linux box performed the best among Opterons, followed by the Newisys and then the Appro, in spite of the Newisys having faster processors and more RAM.
When running the Zeus Web server, the performance of the HP Itanium system was — while still slower — very close to the three Opterons, an indication that Apache on Red Hat Linux was not as optimized for the Itanium as Apache on SuSE Linux was optimized for the Opteron. (With more complex test suites, where each client was making multiple requests including large graphics, the HP fell a little further behind.) The Opterons all had Apache 1.3.26-105 installed, while the HP Itanium ran Apache 1.3.27-2. The table below shows the results running the Zeus Web server under 1,000 virtual clients.
The lesson for anyone looking to squeeze real 64-bit performance from an Opteron or Itanium system is to make sure your applications are not only 64-bit capable, but optimized for 64-bit, as the Zeus Web server is. And of course, those planning to run Web sites on Itanium clearly should do so on Zeus, rather than Apache.
Other conclusions? In general, the tests show the Opteron has a slight edge over Itanium, at least for now. Among all my tests, the greatest difference between the highest-performing Opteron and the Itanium is about 20 percent. With the 1.3GHz and 1.5GHz Itanium 2 processors due by the time you read this, the advantage should be moving to the Itanium. In addition to increasing clock speed, the new chips double the L2 cache size to 6MB, compared to the 3MB in the HP Itanium system I tested.
Regardless of whether you choose Opteron or Itanium for Web serving, you’ll reap huge performance gains over 32-bit Xeon or Athlon. Even Apache on Itanium showed a 40 percent improvement over the performance of the dual Xeon system I used as a baseline.
Web server requests don’t stress an entire system as much as some other tasks. Opteron and Itanium systems should demonstrate even greater performance gains when handling very large databases, for example. Nevertheless, Web server performance is a good general indicator of the kinds of gains possible with 64-bit hardware and 64-bit code.
We should all look forward to the day when the applications we’re running now are available and optimized for Opteron and Itanium.
Overall Score (100%)
|Pogo Linux PerformanceWare 1264||6.0||7.0||7.0||8.0||8.0||6.0|
|Newisys 2100 Server||8.0||8.0||8.0||7.0||6.0||8.0|
|HP Integrity Server rx2600||9.0||6.0||9.0||7.0||8.0||9.0|
Looking for the missing free copy icon? It's been replaced. There's a new direct link that works like a...
Supreme Court's decision is bad news for developers targeting the U.S. market, who will now have to...
The transition from command line to line-of-command requires a new mind-set -- and a thick skin
If an 'independent' code review says a product is totally secure, you aren't hearing the full story
A spate of projects from IBM's DeveloperWorks Open portal covers everything from improving Spark...
Built for development teams, Git can’t meet enterprise scalability and security requirements on its own...
AWS's developer-focused approach is one lesson enterprises should glean from the cloud leader