The Free Lunch Is over: a Fundamental Turn Toward Concurrency in Software
script>
script>
script>
script>
script>
script>
script>
script>
News
January
The New C column: Trip Report: October 2004. This month, an update on the most recent C standards meeting, and the progress on the C 0x standard language and library Conversations column:.. Tagged Unions Some of our C heritage is Slim and fast 梐 nd fragile 梚 n the name of perform. UNIONS Are Such a beast.
February
Sutter's Mill column: The Concurrency Revolution The biggest sea change in software development since the OO revolution is knocking at the door, and its name is Concurrency Conversations column:... Order, Order Getting the order right for a standard associative container can take a Bit of thought.
March
Dr. Dobb's Journal feature: The Free Lunch Is Over:. A Fundamental Turn Toward Concurrency in Software A more in-depth and detailed version of the February CUJ column The Concurrency Revolution (see above).
The Free Lunch Is over: a Fundamental Turn Toward Concurrency in Software
The biggest sea change in software development since the OO revolution is knocking at the door, and its name is Concurrency. This article will appear in Dr. Dobb's Journal, 30 (3), March 2005. A much briefer version under the title "The Concurrency Revolution "will appear in C / C Users Journal, 23 (2), February 2005. 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 Already 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, as many companies showed new or updated multicore processors. Looking back, it Zha 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 Yun l describe the changing face of hardware, why it suddenly does matter to software, and how specifically it matters to you and is going to change the way you will likely be writing software in the future. Arguably, the free lunch has already been over for a year or two, only we Zhen e just now noticing.The Free Performance Lunch
There an interesting phenomenon that Zha Zha known as You ndy 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 ten times as fast, and software will usually find ten times as much to do (or, in some cases, will feel at liberty to do it ten 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 Yang the only measure of performance, or even necessarily a good one, but it Zha an instructive one: We Zhen e used to seeing 500MHz CPUs give way to 1GHz CPUs give way to 2GHz CPUs, and so on Today we Zhen e in the 3GHz range on mainstream computers The key question is:.. When will it end After? ALL, MOORE 抯 LAW Predicti growth, and clearly exponential growth can Yang continue forever before we reach hard physical limits; light isn Yang getting any faster The growth must eventually slow down and even end (Caveat:.. Yes, Moore Zha Law applies principally to transistor densities, but the same kind of exponential growth has occurred in related areas such as clock speeds. There Zha even faster growth in other spaces, most notably the data storage explosion, but that important trend belongs in a different article.) If you Zhen ea software developer, chances are that you have already been riding the Wei ree lunch wave of desktop computer performance Is your application Zha performance borderline for some local operations Huang ot to worry, the conventional (if suspect) wisdom goes?.??;
E omorrow Zha processors will have even more throughput, and anyway today Zha applications are increasingly throttled by factors other than CPU throughput and memory speed (eg, they Zhen e often I / O-bound, network-bound, 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:. clock speed
Execution Optimization
cache
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 Zha 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 / 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.
Chip Designers Are Under So Much Pressure To Deliver Ever-Faster Cpus That They 抣 L Risk Changing The Meaning of Your Program, And Possibly Break It, in Order To Make It Run Faster
Brief aside on instruction reordering and memory models: Note that some of what I just called pluck ptimizations are actually far more than optimizations, in that they can change the meaning 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 Yang hurt a fly, and wouldn Yang 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 Zhen e under so much pressure that they Yun l risk changing the meaning of your program, and possibly break it, in order 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 Zha 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 Yang 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 stockade nd you can Yang 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 and more 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. I 抣 L Talk A Little More About The Importance of Cache Later ON.) 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 will directly lead to speedups in sequential (nonparallel, single-threaded, single-process) applications, as well as applications that do make use of concurrency that Zha important, because the vast majority of today Zha applications are single-threaded, for good reasons that I Yun l get into further below of course, compilers have had to keep up;.. sometimes you need to recompile your application, and target a specific minimum level of CPU, in order to benefit from new instructions (eg, MMX, SSE) and some new CPU features and characteristics. But, by and large, even old applications have always run significantly faster Ren ven 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.Obstacles, and Why You Don Yang Have 10GHz Tod AY
Figure 1: Intel CPU Introductions (Sources: Intel, Wikipedia)
CPU performance growth as we have known it hit a wall two years ago. Most people have only recently started to notice. You can get similar graphs for other chips, but I Bing 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. Around the beginning of 2003, you Yun l note a disturbing sharp turn in the previous Trend Toward EVER-FASTER CPU Clock Speeds. I 抳 E 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 not just one but several physical issues, notably heat (too much of it and too hard to dissipate), power consumption (too high), and current leakage problems Quick:.?? What Zha 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, now in early 2005 we should have the first 10GHz Pentium-family chips. A quick look around shows that, well, actually, we don Yang. What Zha more , SUCH Chips Are Not Even ON THE HORZON 梬 E Have No Good See the About 4GHZ? WE 抮 E AT 3.4GHZ ALREADY 梥 URELY 4GHZ CAN 抰 Be Far Away? Alas Even 4GHz Seems To Be Remote Indeed. in Mid-2004, As You Probly Know, Intel First Delayed ITS Planned Introduction of A 4GHz C hip 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 in early 2005 (already included in Figure 1 as the upper-right-most Dot), But The Clock Race Really Is over, AT Least for Now;
Intel Zha and most processor vendors? Future lies elsewhere as chip companies aggressively pursue the same new multicore directions. We Yun l probably see 4GHz CPUs in our mainstream desktop machines someday, but it won Yang be in 2005. Sure, Intel has samples of their chips running at even higher speeds in the lab Di ut only by heroic efforts, such as attaching hideously impractical quantities of cooling equipment. You won Yang have that kind of cooling hardware in your office any day soon, let alone on your lap while computing on The Plane.TanStaafl: Moore 抯 Law and the next generation (s)
揟 Here ain 抰 No Such Thing as a free lunch.? / I> 桼. A. Heinlein, The Moon Is a Harsh Mistress
Does this mean Moore Zha Law is over? Interestingly, the answer in general seems to be no. Of course, like all exponential progressions, Moore Zha Law must end someday, but it does not seem to be in danger for a few more years yet . 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 Zha Law-like throughput gains for some years to come.
Myths and realities: 2 x 3GHz <6 GHz
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 Yang mean getting two times the performance. Similarly, most multi-threaded applications won Yang run TWICE AS FAST ON A DUAL-CORE BOX. The Yy Should Run Faster Than on a Single-CPU;
the performance gain just isn Yang linear, that Zha 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 Yang 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 Yang be well utilized. (Despite this, I will speculate that today Zha single-threaded applications as actually used in the field could actually 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 OT herwise 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 Zhen e 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 Yang going to be maxing out the CPU anyway so one of the cores will be mostly idle. (Again, The Spyware CAN Share the os 抯 Core Most of the Time.)
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
MultiCore
cache
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 hyper-threaded 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% performance boost for reasonably well- written multi-threaded applications, or even as much as 40% under ideal conditions for carefully written multi-threaded applications. That Zha good, but it Zha hardly double, and it doesn Yang 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 in 2005, vary in their level of integration but are functionally simil ar. AMD Zha seems to have some initial performance design advantages, such as better integration of support functions on the same die, whereas Intel Zha initial entry basically just glues together two Xeons on a single die. The performance gains should initially be about the same AS Having a True Dual-CPU System (ONLY THE SYSTEM WILL BE CHEESN 抰 HAVE TO HAVE TWO SOCKETS AND ASSOCIATED 揼 LuE?
chippery), which means something less than double the speed even in the ideal case, and just like today it will boost reasonably well-written multi-threaded 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 Yang want to touch RAM if you can help it. On today Zha systems, a cache miss that goes out to main memory often costs 10 to 50 times as much 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 Zha working set. Fits Into cache, we Zhen e golden, and if it doesn Yang, we Zhen e 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, grind fine ache is king.? / P> (aside: here 抯 an Anecdote to Demonstrate 搒 PACE IS SPEED? That Recently Hit My Compiler Team. The Compiler Uses The Same Source Base for the 32-bit and 64-bit compilers;
The Code Is Just Compiled AS Either A 32-Bit Process OR A 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 Yang 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 Zha working set. that bigger working set caused a performance penalty that almost exactly offset the code execution performance increase we Yun 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 the hardware mean for the way we write software By now you Ni e probably noticed the basic answer, so let Zha 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) naissance 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 Yang 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 Easier Said Than Done, Because Not All Problems Are INHERENTLY Parallel IZABLE AND BECAUSE CONCURRENT ProgramMing IS Hard. I Can Hear The Howls of Protest: 揅 oncurrency? That 抯 NOT NEWS!
People are already writing concurrent applications.?That Zha 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 OO didn Yang become a revolution, and dominant in the mainstream, until the 1990s. Why then? The reason the revolution happened was primarily that 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 Zha strengths in abstraction and dependency management made it a necessity for achieving large-scale software development that is economical, reliable, and repeatable.Concurrency is the next major revolution in how we write software
Similarly, we Ni e 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 Ni e witnessed incrementally more and more programmers writing concurrent (multi-threaded, multi-process ) 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 reasons that I Yun l summarize in the next section. by the way, on the matter of hype: People have always been quick to announce E he next software development revolution, usually about their own brand-new technology Don Yang 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 nece ssary: You can only base a software development revolution on a technology that Zha 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 Zha solid enough to be broadly usable without performance cliffs and other gotchas. As a result, true software development revolutions like OO happen around technologies that have already been undergoing refinement for years, often decades. Even in Hollywood, most genuine pluck vernight 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 OO 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 Yang 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 I Yun l say more about this later on Probably the greatest cost of concurrency is that concurrency really is hard..:
The programming model, meaning the model in the programmer Zha head that he needs to reason reliably about his program, is much harder than it is for sequential control flow. Everybody who learns concurrency thinks they understand it, ends up finding mysterious races they thought weren Yang possible, and discovers that they didn Yang actually understand it yet after all. As the developer learns 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 Yang get caught in testing, however, except in shops that understand why and how to do real stress testing, is those latent concurrency bugs that surface only on true multiprocessor systems, where the threads aren Yang just being switched around on A Single Processor But WHERE THEY REALLY DO EXECUTE TRULY SIMULTAS OF ERRORS. this is the next jolt for people who though That Surely Now They Know How T o write concurrent code: I Ni e 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 Zha 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 soap oing straight to the least forgiving, truly parallel ENVIRONMENT THINGES You Got Wrong. Even WHEN YOAVE A Team That Can Reliably Write Safe Concurrent Code, There Are Other Pitfalls
for example, concurrent code that is completely safe but isn Yang any faster than it was on a single-core machine, typically because the threads aren Yang independent enough and share a dependency on a single resource which re-serializes the program Zha execution. This STUFF GETS PRETTY SUBTLE.THE VAST Majority of Program Today Don 抰 Grok Concurrency, Just As The Vast Majority of Program 13 YEARS AGO DIDN 抰 YET GROK Objects
Just as it is a leap for a structured programmer to learn OO (what Zha an object? What Zha a virtual function? How should I use inheritance? And beyond the prop hats? And Useful ows,? Why are the correct design practices actually correct ?), it Zha a leap of about the same magnitude for a sequential programmer to learn concurrency (what Zha a race? what Zha a deadlock? how can it come up, and how do I avoid it? what constructs actually serialize the program that I thought was parallel? how is the message queue my friend? and beyond the prop hats? and Useful ows,? why are the correct design practices actually correct?). The vast majority of programmers today don Yang grok concurrency, just as the vast majority of programmers 15 years ago didn Yang yet grok objects. But the concurrent programming model is learnable, particularly if we stick to message- and lock-based programming, and once grokked it isn Yang that much harder than OO 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 above to message- and 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. Frankly, Even Lock-based Programming IS Hazardous.) What IT Means for US
Okay, back to what it means for us. 1. The clear primary consequence we Ni e already covered is that applications will increasingly need to be concurrent if they want to fully exploit CPU throughput gains that have now started becoming available and will continue to materialize . over the next several years For example, Intel is talking about someday producing 100-core chips; a single-threaded application can exploit at most 1/100 of such a chip Zha potential throughput Bian h, performance doesn Yang matter so much,. computers just keep getting faster? has always been a na soughing of wind e statement to be viewed with suspicion, and for the near future it will almost always be simply wrong.Applications will increasingly need to be concurrent if they want to fully exploit continuing exponential CPU throughput Gains Efficiency and Performance Optimization Will Get More, Not Less, Important
NOW, Not All Applications (OR, More Precisely, Important Operations of An Application) Are Amenable To Parallelization. True, Some Problems, Such As CompiLation, Are Almost Ideally Parallelizable. But Others Aren;
the usual counterexample here is that just because it takes one woman nine months to produce a baby doesn Yang imply that nine women could produce one baby in one month. You Ni e probably come across that analogy before. But did you notice the problem with leaving ? the analogy at that Here Zha the trick question to ask the next person who uses it on you: Can you conclude from this that the Human Baby Problem is inherently not amenable to parallelization Usually people relating this analogy err in quickly concluding that it demonstrates? an inherently nonparallel problem, but that Zha actually not necessarily correct at all. It is indeed an inherently nonparallel problem if the goal is to produce one child. It is actually an ideally parallelizable problem if the goal is to produce many children! Knowing the real Goals can make all the difference. This Basic Goal-Oriented Principle Is Something To Keep In Mind When Considering WHETHER AND HOW Parallelize Your Software. 2. Perhaps a Less Obvious Co nsequence is that applications are likely to become increasingly CPU-bound. Of course, not every application operation will be CPU-bound, and even those that will be affected won Yang become CPU-bound overnight if they aren Yang already, but we seem to Have reached the end of the 揳 PPLICAS Are INCRESINGLY I / O-BOUND OR NETWORK-BOT OR DATABASE-BOT? TREND, BECAUSE PERFORMANCE IN THOSE AREAS IS STILL IMPROVING RAPIDLY (Gigabit WiFi, Anyone?
) While traditional CPU performance-enhancing techniques have maxed out Consider:.. We Zhen e stopping in the 3GHz range for now Therefore single-threaded programs are likely not to get much faster any more for now except for benefits from further cache size growth ( which is the main good news). Other gains are likely to be incremental and much smaller than we Ni e been used to seeing in the past, for example as chip designers find new ways to keep pipelines full and avoid stalls, which are areas where the low-hanging fruit has already been harvested. The demand for new application features is unlikely to abate, and even more so the demand to handle vastly growing quantities of application data is unlikely to stop accelerating. As we continue to demand that programs do more , they will increasingly often find that they run out of CPU to do it unless they can code for concurrency. There are two ways to deal with this sea change toward concurrency. One is to redesign your applications for concurr ency, as above The other is to be frugal, by writing code that is more efficient and less wasteful This leads to the third interesting consequence:... 3. Efficiency and performance optimization will get more, not less, important Those languages that already Lend Themselves To Heavy Optimization Will Find New Life;
those that don Yang will need to find ways to compete and become more efficient and optimizable. Expect long-term increased demand for performance-oriented languages and systems. 4. Finally, programming languages and systems will increasingly be forced to deal well with concurrency. The Java language has included support for concurrency since its beginning, although mistakes were made that later had to be corrected over several releases in order to do concurrent programming more correctly and efficiently. The C language has long been used to write heavy-duty multithreaded systems well, but it has no standardized support for concurrency at all (the ISO C standard doesn Yang even mention threads, and does so intentionally), and so typically the concurrency is of necessity accomplished by using nonportable platform-specific concurrency features and libraries. ( IT 抯 Also offen Incomplete; for Example, Static Variables Must Be Initialized Only, Which Typically Requires That The Compiler WRA p them with a lock, but many C implementations do not generate the lock.) Finally, there are a few concurrency standards, including pthreads and OpenMP, and some of these support implicit as well as explicit parallelization. Having the compiler look at your single -threaded program and automatically figure out how to parallelize it implicitly is fine and dandy, but those automatic transformation tools are limited and don Yang yield nearly the gains of explicit concurrency control that you code yourself. The mainstream state of the art revolves around lock- Based Programming, Which is Subtle and Hazardous. We desperately Need a higher-level programming model for concurrency Than Languages Offer; i '
Ll Have More To Say About That Soon.conclusion
If you haven Yang 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 Zha requirements, pitfalls, styles, and idioms. A few rare classes of applications are naturally parallelizable, but most aren Yang. Even when you know exactly where you Zhen e CPU- bound, you may well find it difficult to figure out how to parallelize those operations; all the most reason to start thinking about it now implicitly parallelizing compilers can help a little, but don Yang expect much;. they can Yang 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 entr busy and that one dessert The filet mignon of throughput gains is still on the menu, but now it costs extra Ren xtra 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 concurrency will let them fully exploit the continuing exponential gains in processor throughput.
CopyRight? 2004 Herb Sutter