A Fundamental Turn Toward Concurrency in Software

The face of hardware is changing, impacting the way you'll be writing software in the future.


March 01, 2005
URL:http://drdobbs.com/web-development/a-fundamental-turn-toward-concurrency-in/184405990

Your free lunch will soon be over. What can you do about it? What are you doing about it. The major processor manufacturers and architectures, from Intel and AMD to Sparc and PowerPC, have run out of room with most of their traditional approaches to boosting CPU performance. Instead of driving clock speeds and straight-line instruction throughput ever higher, they are instead turning en masse to hyperthreading and multicore architectures. Both of these features are available on chips today; in particular, multicore is available on current PowerPC and Sparc IV processors, and is coming in 2005 from Intel and AMD. Indeed, the big theme of the 2004 In-Stat/MDR Fall Processor Forum was multicore devices, with many companies showing new or updated multicore processors. Looking back, it's not much of a stretch to call 2004 the year of multicore.

And that puts us at a fundamental turning point in software development, at least for the next few years and for applications targeting general-purpose desktop computers and low-end servers (which happens to account for the vast bulk of the dollar value of software sold today). In this article, I describe the changing face of hardware, why it suddenly does matter to software, and how it specifically matters to you and changes the way you'll likely be writing software in the future.

Arguably, the free lunch has already been over for a year or two, only we're just now noticing.

The Free Performance Lunch

There's an interesting phenomenon known as "Andy giveth, and Bill taketh away." No matter how fast processors get, software consistently finds new ways to eat up the extra speed. Make a CPU 10 times as fast, and software usually finds 10 times as much to do (or in some cases, will feel at liberty to do it 10 times less efficiently). Most classes of applications have enjoyed free and regular performance gains for several decades, even without releasing new versions or doing anything special because the CPU manufacturers (primarily) and memory and disk manufacturers (secondarily) have reliably enabled ever-newer and ever-faster mainstream systems. Clock speed isn't the only measure of performance, or even necessarily a good one, but it's an instructive one: We're used to seeing 500MHz CPUs give way to 1GHz CPUs, which give way to 2GHz CPUs, and so on. Today, we're in the 3GHz range on mainstream computers.

The key question is: When will it end? After all, Moore's Law predicts exponential growth, and clearly exponential growth can't continue forever before we reach hard physical limits; light isn't getting any faster. The growth must eventually slow down and even end. (Caveat: Yes, Moore's Law applies principally to transistor densities, but the same kind of exponential growth has occurred in related areas such as clock speeds. There's even faster growth in other spaces, most notably the data storage explosion, but that important trend belongs in a different article.)

If you're a software developer, chances are you have already been riding the "free lunch" wave of desktop computer performance. Is your application's performance borderline for some local operations? "Not to worry," the conventional (if suspicious) wisdom goes, "tomorrow's processors will have even more throughput, and anyway, today's applications are increasingly throttled by factors other than CPU throughput and memory speed (for instance, they're often I/O-bound, network-bound, or database-bound)." Right?

Right enough, in the past. But dead wrong for the foreseeable future.

The good news is that processors are going to continue to become more powerful. The bad news is that, at least in the short term, the growth will come mostly in directions that do not take most current applications along for their customary free ride.

Over the past 30 years, CPU designers have achieved performance gains in three main areas, the first two of which focus on straight-line execution flow:

Increasing clock speed is about getting more cycles. Running the CPU faster more or less directly means doing the same work faster.

Optimizing execution flow is about doing more work per cycle. Today's CPUs sport some more powerful instructions, and they perform optimizations that range from the pedestrian to the exotic, including pipelining, branch prediction, executing multiple instructions in the same clock cycle(s), and even reordering the instruction stream for out-of-order execution. These techniques are all designed to make the instructions flow better and/or execute faster, and to squeeze the most work out of each clock cycle by reducing latency and maximizing the work accomplished per clock cycle.

Note that some of what I just called "optimizations" are actually far more than optimizations, in that they can change the meanings of programs and cause visible effects that can break reasonable programmer expectations. This is significant. CPU designers are generally sane and well-adjusted folks who normally wouldn't hurt a fly and wouldn't think of hurting your code...normally. But in recent years, they have been willing to pursue aggressive optimizations just to wring yet more speed out of each cycle, even knowing full well that these aggressive rearrangements could endanger the semantics of your code. Is this Mr. Hyde making an appearance? Not at all. That willingness is simply a clear indicator of the extreme pressure the chip designers face to deliver ever-faster CPUs; they're under so much pressure that they'll risk changing the meaning of your program, and possibly break it, to make it run faster. Two noteworthy examples in this respect are write reordering and read reordering: Allowing a processor to reorder write operations has consequences that are so surprising, and break so many programmer expectations, that the feature generally has to be turned off because it's too difficult for programmers to reason correctly about the meaning of their programs in the presence of arbitrary write reordering. Reordering read operations can also yield surprising visible effects, but that is more commonly left enabled anyway because it isn't quite as hard on programmers (and the demands for performance cause designers of operating systems and operating environments to compromise and choose models that place a greater burden on programmers because that is viewed as a lesser evil than giving up the optimization opportunities).

Finally, increasing the size of on-chip cache is about staying away from RAM. Main memory continues to be so much slower than the CPU that it makes sense to put the data closer to the processor—and you can't get much closer than being right on the die. On-die cache sizes have soared, and today most major chip vendors will sell you CPUs that have 2MB of on-board L2 cache. (Of these three major historical approaches to boosting CPU performance, increasing cache is the only one that will continue in the near term.)

Okay. So what does this mean?

A fundamentally important thing to recognize about this list is that all of these areas are concurrency agnostic. Speedups in any of these areas directly lead to speedups in sequential (nonparallel, single-threaded, single-process) applications, as well as applications that do make use of concurrency. That's important because the vast majority of today's applications are single-threaded—and for good reasons.

Of course, compilers have had to keep up; sometimes, you need to recompile your application, and target a specific minimum level of CPU, to benefit from new instructions (MMX, SSE, and the like) and some new CPU features and characteristics. But, by and large, even old applications have always run significantly faster—even without being recompiled to take advantage of all the new instructions and features offered by the latest CPUs.

That world was a nice place to be. Unfortunately, it has already disappeared.

Why You Don't Have 10GHz Today

You can get similar graphs for other chips, but I'm going to use Intel data here. Figure 1 graphs the history of Intel chip introductions by clock speed and number of transistors. The number of transistors continues to climb, at least for now. Clock speed, however, is a different story.

Figure 1: Intel CPU introductions (sources: Intel, Wikipedia).

Around the beginning of 2003, you'll note a disturbing sharp turn in the previous trend toward ever-faster CPU clock speeds. I've added lines to show the limit trends in maximum clock speed; instead of continuing on the previous path, as indicated by the thin dotted line, there is a sharp flattening. It has become harder and harder to exploit higher clock speeds due to several physical issues, notably heat (too much of it and too hard to dissipate), power consumption (too high), and current leakage problems.

In short, CPU performance growth as we have known it hit a wall two years ago. Most people have only recently started to notice.

Quick: What's the clock speed on the CPU(s) in your current workstation? Are you running at 10GHz? On Intel chips, we reached 2GHz a long time ago (August 2001), and according to CPU trends before 2003, we now should have the first 10GHz Pentium-family chips. A quick look around shows that, well, actually, we don't. What's more, such chips are not even on the horizon—we have no good idea at all about when we might see them appear.

Well, then, what about 4GHz? We're at 3.4GHz already—surely 4GHz can't be far away? Alas, even 4GHz seems to be remote indeed. In mid-2004, as you probably know, Intel first delayed its planned introduction of a 4GHz chip until 2005, and then in fall 2004, it officially abandoned its 4GHz plans entirely. As of this writing, Intel is planning to ramp up a little further to 3.73GHz early this year (already included in Figure 1 as the upper-right-most dot), but the clock race really is over, at least for now; Intel's and most processor vendors' futures lie elsewhere as chip companies aggressively pursue the same new multicore directions.

We'll probably see 4GHz CPUs in our mainstream desktop machines someday, but it won't be in 2005. Sure, Intel has samples of their chips running at even higher speeds in the lab—but only by heroic efforts, such as attaching hideously impractical quantities of cooling equipment. You won't have that kind of cooling hardware in your office any day soon, let alone on your lap while computing on the plane.

TANSTAAFL: Moore's Law and The Next Generation(s)

TANSTAAFL=There ain't no such thing as a free lunch.
—R.A. Heinlein,
The Moon Is a Harsh Mistress

Does this mean Moore's Law is over? Interestingly, the answer in general seems to be "no." Of course, like all exponential progressions, Moore's Law must end someday, but it does not seem to be in danger for a few more years. Despite the wall that chip engineers have hit in juicing up raw clock cycles, transistor counts continue to explode, and it seems CPUs will continue to follow Moore's Law-like throughput gains for some years to come.

The key difference, which is the heart of this article, is that the performance gains are going to be accomplished in fundamentally different ways for at least the next couple of processor generations. And most current applications will no longer benefit from the free ride without significant redesign.

For the near-term future, meaning for the next few years, the performance gains in new chips will be fueled by three main approaches, only one of which is the same as in the past. The near-term future performance growth drivers are:

Hyperthreading is about running two or more threads in parallel inside a single CPU. Hyperthreaded CPUs are already available today, and they do allow some instructions to run in parallel. A limiting factor, however, is that although a hyperthreaded CPU has some extra hardware (including extra registers), it still has just one cache, one integer math unit, one FPU, and in general, just one each of most basic CPU features. Hyperthreading is sometimes cited as offering a 5 to 15 percent performance boost for reasonably well-written multithreaded applications, or even as much as 40 percent under ideal conditions for carefully written multithreaded applications. That's good, but it's hardly double, and it doesn't help single-threaded applications.

Multicore is about running two or more actual CPUs on one chip. Some chips, including Sparc and PowerPC, have multicore versions available already. The initial Intel and AMD designs, both due this year, vary in their level of integration but are functionally similar. AMD's seems to have some initial performance design advantages, such as better integration of support functions on the same die; whereas Intel's initial entry basically just glues together two Xeons on a single die. The performance gains should initially be about the same as having a dual-CPU system (only the system will be cheaper because the motherboard doesn't have to have two sockets and associated "glue" chippery), which means something less than double the speed even in the ideal case. Just like today, it will boost reasonably well-written multithreaded applications—not single-threaded ones.

Finally, on-die cache sizes can be expected to continue to grow, at least in the near term. Of these three areas, only this one will broadly benefit most existing applications. The continuing growth in on-die cache sizes is an incredibly important and highly applicable benefit for many applications, simply because space is speed. Accessing main memory is expensive, and you really don't want to touch RAM if you can help it. On today's systems, a cache miss that goes out to main memory typically costs about 10 to 50 times as much as getting the information from the cache; this, incidentally, continues to surprise people because we all think of memory as fast, and it is fast compared to disks and networks, but not compared to on-board cache, which runs at faster speeds. If an application's working set fits into cache, we're golden; and if it doesn't, we're not. That is why increased cache sizes will save some existing applications and breathe life into them for a few more years without requiring significant redesign: As existing applications manipulate more and more data, and as they are incrementally updated to include more code for new features, performance-sensitive operations need to continue to fit into cache. As the Depression-era old-timers will be quick to remind you, "Cache is king."

(Aside: Here's an anecdote to demonstrate "space is speed" that recently hit my compiler team. The compiler uses the same source base for 32-bit and 64-bit compilers; the code is just compiled as either a 32-bit process or a 64-bit one. The 64-bit compiler gained a great deal of baseline performance by running on a 64-bit CPU, principally because the 64-bit CPU had many more registers to work with and had other code performance features. All well and good. But what about data? Going to 64 bits didn't change the size of most of the data in memory, except that (of course) pointers in particular were now twice the size they were before. As it happens, our compiler uses pointers much more heavily in its internal data structures than most other kinds of applications ever would. Because pointers were now 8 bytes instead of 4 bytes, a pure data size increase, we saw a significant increase in the 64-bit compiler's working set. That bigger working set caused a performance penalty that almost exactly offset the code execution performance increase we'd gained from going to the faster processor with more registers. As of this writing, the 64-bit compiler runs at the same speed as the 32-bit compiler, even though the source base is the same for both and the 64-bit processor offers better raw processing throughput. Space is speed.)

But cache is it. Hyperthreading and multicore CPUs will have nearly no impact on most current applications.

So what does this change in hardware mean for the way we write software? By now, you've probably noticed the basic answer, so let's consider it and its consequences.

What This Means for Software: The Next Revolution

In the 1990s, we learned to grok objects. The revolution in mainstream software development from structured programming to object-oriented programming was the greatest such change in the past 20 years, and arguably in the past 30 years. There have been other changes, including the most recent (and genuinely interesting) nascence of web services, but nothing that most of us have seen during our careers has been as fundamental and as far reaching a change in the way we write software as the object revolution.

Until now. Starting today, the performance lunch isn't free any more. Sure, there will continue to be generally applicable performance gains that everyone can pick up, thanks mainly to cache size improvements. But if you want your application to benefit from the continued exponential throughput advances in new processors, it will need to be a well-written, concurrent (usually multithreaded) application. And that's easier said than done, because not all problems are inherently parallelizable and because concurrent programming is hard.

I can hear the howls of protest: "Concurrency? That's not news! People are already writing concurrent applications." That's true. Of a small fraction of developers.

Remember that people have been doing object-oriented programming since at least the days of Simula in the late 1960s. But OOP didn't become a revolution, and dominant in the mainstream, until the 1990s. Why then? The reason the revolution happened was primarily because our industry was driven by requirements to write larger and larger systems that solved larger and larger problems and exploited the greater and greater CPU and storage resources that were becoming available. OOP's strengths in abstraction and dependency management made it a necessity for achieving large-scale software development that is economical, reliable, and repeatable.

Similarly, we've been doing concurrent programming since those same dark ages, writing coroutines and monitors and similar jazzy stuff. And for the past decade or so, we've witnessed incrementally more and more programmers writing concurrent (multithreaded, multiprocess) systems. But an actual revolution marked by a major turning point toward concurrency has been slow to materialize. Today, the vast majority of applications are single-threaded, and for good reason.

By the way, on the matter of hype: People have always been quick to announce "the next software development revolution," usually about their own brand-new technology. Don't believe it. New technologies are often genuinely interesting and sometimes beneficial, but the biggest revolutions in the way we write software generally come from technologies that have already been around for some years and have already experienced gradual growth before they transition to explosive growth. This is necessary: You can only base a software development revolution on a technology that's mature enough to build on (including having solid vendor and tool support), and it generally takes any new software technology at least seven years before it's solid enough to be broadly usable without performance cliffs and other gotchas. As a result, true software development revolutions like OOP happen around technologies that have already been undergoing refinement for years, often decades. Even in Hollywood, most genuine "overnight successes" have really been performing for many years before their big break.

Concurrency is the next major revolution in how we write software. Different experts still have different opinions on whether it will be bigger than OO, but that kind of conversation is best left to pundits. For technologists, the interesting thing is that concurrency is of the same order as OOP both in the (expected) scale of the revolution and in the complexity and learning curve of the technology.

Benefits and Costs of Concurrency

There are two major reasons for which concurrency, especially multithreading, is already used in mainstream software. The first is to logically separate naturally independent control flows; for example, in a database replication server I designed, it was natural to put each replication session on its own thread because each session worked completely independently of any others that might be active (as long as they weren't working on the same database row). The second and less common reason to write concurrent code in the past has been for performance, either to scalably take advantage of multiple physical CPUs or to easily take advantage of latency in other parts of the application; in my database replication server, this factor applied as well, and the separate threads were able to scale well on multiple CPUs as our server handled more and more concurrent replication sessions with many other servers.

There are, however, real costs to concurrency.

Some of the obvious costs are actually relatively unimportant. For example, yes, locks can be expensive to acquire, but when used judiciously and properly, you gain much more from the concurrent execution than you lose on the synchronization, if you can find a sensible way to parallelize the operation and minimize or eliminate shared state.

Perhaps the second-greatest cost of concurrency is that not all applications are amenable to parallelization.

Probably the greatest cost of concurrency is that concurrency really is hard: The programming model, meaning the model in programmers' heads that they need to reason reliably about their program, is much harder than it is for sequential control flow.

Everybody who learns concurrency thinks he understands it, but ends up finding mysterious races he thought weren't possible and discovers that he didn't actually understand it after all. As developers learn to reason about concurrency, they find that usually those races can be caught by reasonable in-house testing, and they reach a new plateau of knowledge and comfort. What usually doesn't get caught in testing, however, except in shops that understand why and how to do real stress testing, are those latent concurrency bugs that surface only on true multiprocessor systems, where the threads aren't just being switched around on a single processor but really do execute truly simultaneously and thus expose new classes of errors. This is the next jolt for developers who thought that, surely now, they know how to write concurrent code: I've come across many teams whose application worked fine even under heavy and extended stress testing, and ran perfectly at many customer sites, until the day that a customer actually had a real multiprocessor machine—and then deeply mysterious races and corruptions started to manifest intermittently. In the context of today's CPU landscape, then, redesigning your application to run multithreaded on a multicore machine is a little like learning to swim by jumping into the deep end—going straight to the least forgiving, truly parallel environment that is most likely to expose the things you got wrong. Even when you have a team that can reliably write safe concurrent code, there are other pitfalls; for example, concurrent code that is completely safe but isn't any faster than it was on a single-core machine, typically because the threads aren't independent enough and share a dependency on a single resource that reserializes the program's execution. This stuff gets pretty subtle.

Just as it is a leap for a structured programmer to learn OOP ("what's an object?" "what's a virtual function?" "how should I use inheritance?" and beyond the "whats" and "hows," "why are the correct design practices actually correct?"), it's a leap of about the same magnitude for a sequential programmer to learn concurrency ("what's a race?" "what's a deadlock?" "how can it come up, and how do I avoid it?" "what constructs actually serialize the program that I thought was parallel?" and beyond the "whats" and "hows," "why are the correct design practices actually correct?"). The vast majority of programmers aren't there today, just as the vast majority of programmers 15 years ago didn't yet grok objects. But the concurrent programming model is learnable, particularly if we stick to lock-based programming, and once grokked, it isn't that much harder than OOP and hopefully can become just as natural. Just be ready and allow for the investment in training and time, for you and for your team.

(I deliberately limit the aforementioned discussion to lock-based concurrent programming models. There is also lock-free programming, supported most directly at the language level in Java 5 and in at least one popular C++ compiler. But concurrent lock-free programming is known to be very much harder for programmers to understand and reason about than even concurrent lock-based programming. Most of the time, only systems and library writers should have to understand lock-free programming, although virtually everybody should be able to take advantage of the lock-free systems and libraries those people produce.)

What it Means for Us

Okay, back to what it means for us.

There are two ways to deal with this sea change toward concurrency. One is to redesign your applications for concurrency. The other is frugality, or writing code that is more efficient and less wasteful. This leads to the third interesting consequence:

Conclusion

If you haven't done so already, now is the time to take a hard look at the design of your application, determine what operations are CPU-sensitive now or are likely to become so soon, and identify how those places could benefit from concurrency. Now is also the time for you and your team to grok concurrent programming's requirements, pitfalls, styles, and idioms.

A few rare classes of applications are naturally parallelizable, but most aren't. Even when you know exactly where you're CPU-bound, you may well find it difficult to figure out how to parallelize those operations; all the more reason to start thinking about it now. Implicitly parallelizing compilers can help a little, but don't expect much; they can't do nearly as good a job of parallelizing your sequential program as you could do by turning it into an explicitly parallel and threaded version.

Thanks to continued cache growth and probably a few more incremental straight-line control flow optimizations, the free lunch will continue a little while longer; but starting today, the buffet will only be serving that one entrée and that one dessert. The filet mignon of throughput gains is still on the menu, but now it costs extra-extra development effort, extra code complexity, and extra testing effort. The good news is that for many classes of applications the extra effort will be worthwhile because it will let them fully exploit the continuing exponential gains in processor throughput.

Myths and Realities: 2×3GHz6GHz

So, a dual-core CPU that combines two 3GHz cores practically offers 6GHz of processing power. Right? Wrong. Even having two threads running on two physical processors doesn't mean getting two times the performance. True, there are some kinds of problems that are inherently parallelizable and can approach linear throughput gains; a typical example is compilation, which can run close to twice as fast on a carefully managed dual-CPU dual-disk system.

Similarly, most multi-threaded applications won't run twice as fast on a dual-core box, although they should run faster than on a single-core CPU. The performance gain just isn't linear, that's all.

Why not? First, there is coordination overhead between the cores to ensure cache coherency (a consistent view of cache and of main memory) and to perform other handshaking. Today, a two- or four-processor machine isn't really two or four times as fast as a single CPU even for multi-threaded applications. The problem remains essentially the same even when the CPUs in question sit on the same die.

Second, unless the two cores are running different processes, or different threads of a single process that are well-written to run independently and almost never wait for each other, they won't be well utilized. (Despite this, I will speculate that today's single-threaded applications as actually used in the field could see a performance boost for most users by going to a dual-core chip, not because the extra core is actually doing anything useful, but because it is running the adware and spyware that infest many users' systems and are otherwise slowing down the single CPU that user has today. I leave it up to you to decide whether adding a CPU to run your spyware is the best solution to that problem.)

If you're running a single-threaded application, then the application can only make use of one core. There should be some speedup as the operating system and the application can run on separate cores, but typically, the OS isn't going to be maxing out the CPU anyway, so one of the cores will be mostly idle. (Again, the spyware can share the OS's core most of the time.)

—H.S.


Herb Sutter chairs the ISO C++ Standards committee and is an architect in Microsoft's Developer Division. His most recent books are Exceptional C++ Style and C++ Coding Standards (Addison-Wesley).

Terms of Service | Privacy Statement | Copyright © 2024 UBM Tech, All rights reserved.