Home > Articles > Programming > C/C++

  • Print
  • + Share This
This chapter is from the book

This chapter is from the book

13.2 A Brief History of Data Sharing

One aspect of the shift happening in computing is the suddenness with which processing and concurrency models are changing today, particularly in comparison and contrast to the pace of development of programming languages and paradigms. It takes years and decades for programming languages and their associated styles to become imprinted into a community's lore, whereas changes in concurrency matters turned a definite exponential elbow starting around the beginning of the 2000s.

For example, our yesteryear understanding of general concurrency1 was centered around time sharing, which in turn originated with the mainframes of the 1960s. Back then, CPU time was so expensive, it made sense to share the CPU across multiple programs controlled from multiple consoles so as to increase overall utilization. A process was and is defined as the state and the resources of a running program. To implement time sharing, the CPU uses a timer interrupt in conjunction with a software scheduler. Upon each timer interrupt, the scheduler decides which process gets CPU time for the next time quantum, thus giving the illusion that several processes are running simultaneously, when in fact they all use the same CPU.

To prevent buggy processes from stomping over one another and over operating system code, hardware memory protection has been introduced. In today's systems, memory protection is combined with memory virtualization to ensure robust process isolation: each process thinks it "owns" the machine's memory, whereas in fact a translation layer from logical addresses (as the process sees memory) to physical addresses (as the machine accesses memory) intermediates all interaction of processes with memory and isolates processes from one another. The good news is that runaway processes can harm only themselves, but not other processes or the operating system kernel. The less good news is that upon each task switching, a potentially expensive swapping of address translation paraphernalia also has to occur, not to mention that every just-switched-to process wakes up with cache amnesia as the global shared cache was most likely used by other processes. And that's how threads were born.

A thread is a process without associated address translation information—a bare execution context: processor state plus stack. Several threads share the address space of a process, which means that threads are relatively cheap to start and switch among, and also that they can easily and cheaply share data with each other. Sharing memory across threads running against one CPU is as straightforward as possible—one thread writes, another reads. With time sharing, the order in which data is written by one thread is naturally the same as the order in which those writes are seen by others. Maintaining higher-level data invariants is ensured by using interlocking mechanisms such as critical sections protected by synchronization primitives (such as semaphores and mutexes). Through the late twentieth century, a large body of knowledge, folklore, and anecdotes has grown around what could be called "classic" multithreaded programming, characterized by shared address space, simple rules for memory effect visibility, and mutex-driven synchronization. Other models of concurrency existed, but classic multithreading was the most used on mainstream hardware.

Today's mainstream imperative languages such as C, C++, Java, or C# have been developed during the classic multithreading age—the good old days of simple memory architectures, straightforward data sharing, and well-understood interlocking primitives. Naturally, languages modeled the realities of that hardware by accommodating threads that all share the same memory. After all, the very definition of multithreading entails that all threads share the same address space, unlike operating system processes. In addition, message-passing APIs (such as the MPI specification [29]) have been available in library form, initially for high-end hardware such as (super)computer clusters.

During the same historical period, the then-nascent functional languages adopted a principled position based on mathematical purity: we're not interested in modeling hardware, they said, but we'd like to model math. And math for the most part does not have mutation and is time-invariant, which makes it an ideal candidate for parallelization. (Imagine the moment when those first mathematicians-turned-programmers heard about concurrency—they must have slapped their foreheads: "Wait a minute!...") It was well noted in functional programming circles that such a computational model does inherently favor out-of-order, concurrent execution, but that potential was more of a latent energy than a realized goal until recent times.

Finally, Erlang was developed starting in the late 1980s as a domain-specific embedded language for telephony applications. The domain required tens of thousands of simultaneous programs running on the same machine and strongly favored a message-passing, "fire-and-forget" communication style. Although mainstream hardware and operating systems were not optimized for such workloads, Erlang initially ran on specialized hardware. The result was a language that originally combined an impure functional style with heavy concurrency abilities and a staunch message-passing, no-sharing approach to communication.

Fast-forward to the 2010s. Today, even run-of-the-mill machines have more than one processor, and the decade's main challenge is to stick ever more CPUs on a chip. This has had a number of consequences, the most important being the demise of seamless shared memory.

One time-shared CPU has one memory subsystem attached to it—with buffers, several levels of caches, the works. No matter how the CPU is time-shared, reads and writes go through the same pipeline; as such, a coherent view of memory is maintained across all threads. In contrast, multiple interconnected CPUs cannot afford to share the cache subsystem: such a cache would need multiport access (expensive and poorly scalable) and would be difficult to place in the proximity of all CPUs simultaneously. Therefore, today's CPUs, almost without exception, come with their own dedicated cache memory. The hardware and protocols connecting the CPU + cache combos together are a crucial factor influencing multiprocessor system performance.

The existence of multiple caches makes data sharing across threads devilishly difficult. Now reads and writes in different threads may hit different caches, so sharing data from one thread to another is not straightforward anymore and, in fact, becomes a message passing of sorts:2 for any such sharing, a sort of handshake must occur among cache subsystems to ensure that shared data makes it from the latest writer to the reader and also to the main memory.

As if things weren't interesting enough already, cache synchronization protocols add one more twist to the plot: they manipulate data in blocks, not individual word reads and word writes. This means that communicating processors "forget" the exact order in which data was written, leading to paradoxical behavior that apparently defies causality and common sense: one thread writes x and then y and for a while another thread sees the new y but only the old x. Such causality violations are extremely difficult to integrate within the general model of classic multithreading, which is imbued with the intuition of time slicing and with a simple memory model. Even the most expert programmers in classic multithreading find it unbelievably difficult to adapt their programming styles and patterns to the new memory architectures.

To illustrate the rapid changes in today's concurrency world and also the heavy influence of data sharing on languages' approach to concurrency, consider the following piece of advice given in the 2001 edition of the excellent book Effective Java [8, Item 51, page 204]:

  • When multiple threads are runnable, the thread scheduler determines which threads get to run and for how long.... The best way to write a robust, responsive, portable multithreaded application is to ensure that there are few runnable threads at any given time.

One startling detail for today's observer is that single-processor, time-sliced threading is not only addressed by the quote above, but actually assumed without being stated. Naturally, the book's 2008 edition3 [9] changes the advice to "ensure that the average number of runnable threads is not significantly greater than the number of processors." Interestingly, even that advice, although it looks reasonable, makes a couple of unstated assumptions: one, that there will be high data contention between threads, which in turn causes degradation of performance due to interlocking overheads; and two, that the number of processors does not vary dramatically across machines that may execute the program. As such, the advice is contrary to that given, repeatedly and in the strongest terms, in the Programming Erlang book [5, Chapter 20, page 363]:

  • Use Lots of Processes This is important—we have to keep the CPUs busy. All the CPUs must be busy all the time. The easiest way to achieve this is to have lots of processes.4 When I say lots of processes, I mean lots in relation to the number of CPUs. If we have lots of processes, then we won't need to worry about keeping the CPUs busy.

Which recommendation is correct? As usual, it all depends. The first recommendation works well on 2001-vintage hardware; the second works well in scenarios of intensive data sharing and consequently high contention; and the third works best in low-contention, high-CPU-count scenarios.

Because of the increasing difficulty of sharing memory, today's trends make data sharing tenuous and favor functional and message-passing approaches. Not incidentally, recent years have witnessed an increased interest in Erlang and other functional languages for concurrent applications.

  • + Share This
  • 🔖 Save To Your Account