Linux and the Alpha
Ever since its announcement in the Fall of 1991, the Alpha architecture (see Reference 3) has been the foundation of the world's fastest systems. In fact, except for one or two brief blips, Alpha systems have been the highest-performing systems based on single-CPU SPECmark performance. With this outstanding performance record comes marketing hype and, sometimes, unrealistic expectations. It is not all that uncommon to find e-mail messages or USENET articles saying things like: “I heard the Alpha is so fast, but now I find that my dusty deck is just 10% faster on the Alpha than on the other system.” So what's the truth? The honest answer is that it depends on what you're doing. Alpha systems are without a doubt fast machines, but it is unreasonable to expect that taking a dusty deck and running it on an Alpha will result in the best possible performance. This is particularly true for programs that were written with the mind-set of the eighties, when CPU cycles were at a premium and memory bandwidth was abundant. Reality looks quite different today: CPU clock-rates above 150MHz are the rule and even laptops can run at 200MHz or more. The result is that, today, the memory system—and not the CPU—is often the first-order bottleneck.
In part 2 of this article, we will demonstrate a few simple techniques that help avoid the memory system bottleneck. Except for one case, the focus is on integer-intensive applications. The topic of optimizing floating-point intensive applications is certainly just as important but, unfortunately, well beyond the scope of this series. The techniques presented can result in tremendous performance improvements. While the techniques will be helpful for all modern systems, they normally extract the biggest benefits on Alpha-based machines. There are a couple of reasons for this bias.
One, the Alpha architecture has been designed with longevity in mind. Specifically, the Alpha architecture should be good for the next 15-25 years, which corresponds roughly to a 1000-fold increase in overall performance. For this reason, some design-tradeoffs were made in favor of long-term viability rather than short-term benefits. For example, the Alpha was right from the start a 64-bit architecture, even though, at the time of its announcement, 32-bit address spaces were considered comfortably large.
Two, the current Alpha implementations are designed to achieve high performance by pushing clock frequency to the limit. This means the CPU-to-memory-system performance gap is the largest for Alpha-based systems. For example, suppose a memory access takes 100ns. On a 500MHz Alpha CPU, this corresponds to 50 clock cycles. In contrast, on a 250MHz CPU, this is only 25 cycles. So the relative performance penalty of accessing memory is much higher on a CPU with higher clock speeds. This may sound like a bad thing, but since the absolute performance is the same, what this really means is that a fast-clock CPU system that is running a memory-bound application will be about as fast as a slower-clock system, but when running a memory-wise application, it will be much faster.
In this part of the series, I present a brief overview of existing and upcoming Alpha implementations. While it is not usually necessary to optimize for a specific CPU, it is helpful to know what the characteristics of current CPUs and systems are. I also discuss a couple of simple performance analysis tools that are available under Linux. When porting legacy code to modern systems, such tools are invaluable, since they avoid wasting time trying to optimize rarely executed code.
So far, the Alpha CPU family tree spans three generations; it all began with the 21064 chip. At the time of its introduction, it was the highest performing CPU, and it still makes for a nice workstation, though it's no longer competitive with the latest generation CPUs. This chip branched off into a version that was called the “Low-Cost Alpha” (LCA), also known as 21066 or 21068. The chip core was identical to the 21064 but it had an integrated memory and PCI-bus controller. This high integration made it possible to build Alpha-based systems at relatively low cost and for the embedded systems market. Unfortunately, the design had a major weakness—the memory system was seriously under-powered. This created the paradoxical situation in which a system based on this chip performed on some applications on average, no better than a 100MHz Pentium, but outperformed a P6 running at 200MHz. As a result, the reaction to this chip varied greatly, and probably resulted in quite a few disappointed customers for Digital. On the other hand, there is no doubt that the low-cost at which 21066-based systems eventually were sold caused a quantum leap in the number of Linux/Alpha users.
Around June 1994, the 21164 chip was announced. It had dramatically improved performance over the 21064 and was the first, and so far only, Alpha CPU to feature a three-level cache hierarchy. The first and second-level caches were both on-chip and only the third-level cache was on the motherboard. This chip, in slightly improved versions, is still going strong. At the Fall 1996 Comdex in Las Vegas, such a chip, coupled with a liquid cooling system, was demonstrated running at 767MHz. Another version, called 21164PC, is scheduled to become available around Spring 1997. It omits the relatively expensive second-level, on-chip cache but adds multi-media extensions and other performance-enhancing features. As the name indicates, this chip is designed to be price-competitive with PC processors, specifically the forthcoming Intel Klamath (an improved P6). While price-competitive, the 21164PC is supposed to deliver over 50% better performance than the Klamath. For this second-generation, low-cost Alpha implementation, it certainly looks like Digital and its co-designer Mitsubishi are not going to repeat the mistakes of the past. The 21164PC promises to be cheap and fast.
If you happen to have a deep pocket or want to take a glance at what PC processors might look like in two or three years, the 21264 might be of interest. It is scheduled to become available in high-end machine during the second half of 1997. With this chip, CPU performance is expected to take another giant leap. Current estimates call for a performance level that is three to four times faster than the fastest CPUs available today.
Between each major chip generation, there are typically “half-generation” CPUs which have improvements that derive primarily from a shrink of the chip manufacturing process. For example, the 21064 chip was followed by the 21064A, and similarly, the 21164 was followed by the 21164A. In the former case, the core of the chip remained virtually identical to the 21064, but the primary caches doubled in size from 8KB to 16KB. In the latter case, instructions for byte and word accesses were added and the maximum clock frequency increased from 333 to 500MHz.
A summary of the performance attributes of the current Alpha chip family is presented in Table 1.
2394s1.ps (min=50, max=52)
Webinar: 8 Signs You’re Beyond Cron
11am CDT, April 29th
Join Linux Journal and Pat Cameron, Director of Automation Technology at HelpSystems, as they discuss the eight primary advantages of moving beyond cron job scheduling. In this webinar, you’ll learn about integrating cron with an enterprise scheduler.Join us!
|Android Candy: Intercoms||Apr 23, 2015|
|"No Reboot" Kernel Patching - And Why You Should Care||Apr 22, 2015|
|Return of the Mac||Apr 20, 2015|
|DevOps: Better Than the Sum of Its Parts||Apr 20, 2015|
|Play for Me, Jarvis||Apr 16, 2015|
|Drupageddon: SQL Injection, Database Abstraction and Hundreds of Thousands of Web Sites||Apr 15, 2015|
- Tips for Optimizing Linux Memory Usage
- "No Reboot" Kernel Patching - And Why You Should Care
- DevOps: Better Than the Sum of Its Parts
- Return of the Mac
- Android Candy: Intercoms
- Drupageddon: SQL Injection, Database Abstraction and Hundreds of Thousands of Web Sites
- Non-Linux FOSS: .NET?
- Play for Me, Jarvis
- Designing Foils with XFLR5