In Vulkan, it is recommended to break the API calls into separate threads for better throughput. I am unsure which category of calls are the computationally expensive one which would cause the thread to block, and thus should be used asynchronously.
As I see it, these are the potential calls/family-of-calls that could take a long time to execute.
vkAcquireImageKHR()
vkQueueSubmit()
vkQueuePresentKHR()
memcpy into mapped memory
vkBegin/EndCommandBuffer
vkCmd* calls for drawing and compute
But, the more I think about them, the more it seems that most would be fairly cheap to call. I'll explain my rational, which is probably flawed.
vkAcquireImageKHR()
This could block, if you choose a timeout. But, it's likely that a sufficiently optimized app would call this function with a 0 timeout, and just do other work if the image is not yet available. So, this function can be made instant. There's no need to wait, if the app is smart enough.
vkQueueSubmit()
This function takes a fence, which will be signaled when the GPU has finished executing the command buffers. So, it doesn't actually wait around while the GPU performs the work. I'm assuming this function is the one that starts the physical movement of the command buffer data to the GPU, but I'm assuming that it tell the hardware to read from some memory location, and then the function returns as quickly as possible. So, it wouldn't wait around while the command buffers get sent to the GPU.
vkQueuePresentKHR()
Signal to the GPU to send some image to the window/monitor. It doesn't have to wait for much, does it?
memcpy into mapped memory
This is probably slow.
vkCmd* calls
This family of calls is the one I'm most unsure about. When I read about threads and Vulkan, it's usually these calls that get put onto the threads. But, what are these calls doing, really? Are they building some opcode buffer, made up of some ints and pointers, to be sent to the GPU? If so, that should be extremely fast. The actual work would be carrying out the operations described by those opcodes.
Define "block". The traditional definition of "block"ing is to wait on some internal synchronization, and thereby taking longer than would strictly be necessary for the operation. Doing a memcpy is not doing any synchronization; it's just copying data.
So you don't seem to be concerned about "block"ing; you're merely talking about what operations are expensive.
vkQueueSubmit does not block. But that doesn't mean it's not expensive. It is not "tell[ing] the hardware to read from some memory location" Just look at its interface. It doesn't take a single command buffer; it takes an arbitrary number of them, which are grouped into batches, with each batch waiting on semaphores before execution, signaling semaphores after execution, and the whole operation signaling a fence.
You cannot reasonably expect an implementation of such a thing to merely copy some pointers around.
And that doesn't even get into issues of different types of command buffers. Submitting SIMULTANEOUS_USE command buffers may require creating temporary copies of its buffered data, so that different batches can contain the same command buffer.
Now obviously, vkQueueSubmit is going to return well before any of the work it submits actually gets executed. But don't make the mistake of thinking that it's free to ship work off to the GPU. The Vulkan specification takes time out in a note to directly tell you not to call the function any more frequently than you can get away with:
Submission can be a high overhead operation, and applications should attempt to batch work together into as few calls to vkQueueSubmit as possible.
The reason to present on the same thread that submitted the CBs that generates the image being presented is not because any of those operations are necessarily slow. It's for simple pragmatism; these three operations (acquire, submit, present) must happen in order. And the simplest and easiest way to ensure that is to do them on the same thread.
You cannot submit work that renders to a swapchain image until you have acquired it. Therefore, either you do it on the same thread, or you have to have some inter-thread communication pipe to tell the thread waiting to build the primary CB what the acquired image is. The two processes cannot overlap.
Unlike acquire, present is a queue operation. And both vkQueueSubmit and vkQueuePresent require that access to their VkQueue parameters must be "externally synchoronized". That of course means that you cannot call them both from different threads, on the same VkQueue, at the same time. So if you tried to do these in parallel, you'd need a mutex or something to synchronize CPU access to the VkQueue.
Whereas if you do them on the same thread, there's no need.
Additionally, in order to present an image, you must provide a semaphore that the present will wait on. This semaphore will get signaled by the batch that generates data for the image. Vulkan requires semaphore signal/wait pairs to be ordered; you cannot perform a queue operation that waits on a semaphore until the operation that signals that semaphore has been submitted. Therefore, either you do it on the same thread in sequence, or you use some inter-thread communication pipe to tell whatever thread is waiting to present the image that the submit operation that renders to it has been issued.
So what is to be gained by splitting these operations up onto different threads? They have to happen in sequence, so you may as well do them in sequence the easiest way that exists: on the same thread.
While timeline semaphores now allow you to call the present function before submitting the work that increments the semaphore counter, you still can't call them on separate threads (without synchronization) because they affect the same queue. So you may as well issue them on the same thread (though not necessarily in acquire, submit, present order).
Ultimately, it's not clear what the point of this exercise is. Yes, an individual vkCmd* call will be pretty fast. So what? In a real scene, you will be calling these functions thousands of times per frame. Spreading them evenly across 4 cores saves you ~4x the performance.
Related
I have read that we use semaphores inside the linux kerenl,and i have read that semaphores has advantages even in one single cpu (we can run only one process\thread). Can anyone please give me an example of a problem that semaphore solves(inside the kernel)?
In my view, there can be a problem only if we have more than one cpu, because two process may call system calls that use the same data structure, and probablly cause problems.
Thank you for your help!
You don't really need more than one CPU for concurrency. The multiple CPUs are really "an implementation detail," a piece of hardware quirkiness that you can abstract away from. Concurrency is a logical property of programs. You can have concurrency without multiple CPUs, and use multiple CPUs without "real concurrency".
Consider a web server. It has to be "concurrent," in the sense that it must serve multiple clients at once, hold information about multiple connections and once, and process multiple requests at once. You can have it literally do this, by having multiple CPUs all working at the same time. Yet, the program only has to appear to do multiple things at once. It could just as well be running on one CPU and context switching to fairly service all the work put to it. The fact that a web-server does multiple things at once is part of its interface: the I/O for the connections are interleaved, if a request has exclusively locked a resource, another request won't start trying to manipulate that same resource, etc. Writing a web server without concurrency produces a program that is wrong.
Semaphores help you with concurrency, by letting you control the way processes access resources. You asked, if you had one process running, how another could run at the same time with only a single core. Well, as I said, concurrency doesn't need multiple cores. The first process can be paused, and the second one started while the first one is still unfinished. This is just an implementation detail; logically, to the program writer, the two processes are running simultaneously, whether there are multiple cores or not. If the program was written without semaphores (or had broken concurrency in some other way), it would be wrong, even on a single core. Physically, this will be because context switching can abruptly pause one computation and start another at any time, and, without semaphores, the newly live thread won't know what resources it can and cannot access. Logically, this will be because the processes are running simultaneously, once you abstract yourself away from the implementation, and, in general, processes running simultaneously can walk over each other if not properly synchronized.
For an example applicable to an OS kernel, consider that every process is logically running concurrently with every other process. A kernel provides the implementation that makes this concurrency work. A resource that two processes may want simultaneously is a hard drive. A semaphore might be used in the kernel to track whether a given drive is currently busy with a read or write. A process trying to read or write to the same disk will ask the kernel to do so, and the kernel can check the semaphore to see that the disk is still busy and force the offending process to wait. Now, an operating system does count as low level code, so in some places, yes, you might want to omit some otherwise vital concurrency safeguards when running on a single CPU, because your job is to handle such implementation details, but higher level parts may still use them.
In contrast, consider a number-crunching program. Let's say it's processing each element of a huge array of data into an equal-sized array of modified data (a functional map operation). It can use multiple CPUs to do this more quickly, but it can also work one CPU. The observable behavior of the program is the same, and you never get any idea that it's doing multiple things at once from its behavior. Numbers go in, numbers come out, who cares what happens in the middle? Writing such a program without the ability to do multiple things at once does not produce a logically incorrect program, just a slow one. Such a program probably does not need semaphores when running on a single CPU, because it didn't need concurrency in the first place.
In my application it is imperative that "state" and "graphics" are processed in separate threads. So for example, the "state" thread is only concerned with updating object positions, and the "graphics" thread is only concerned with graphically outputting the current state.
For simplicity, let's say that the entirety of the state data is contained within a single VkBuffer. The "state" thread creates a Compute Pipeline with a Storage Buffer backed by the VkBuffer, and periodically vkCmdDispatchs to update the VkBuffer.
Concurrently, the "graphics" thread creates a Graphics Pipeline with a Uniform Buffer backed by the same VkBuffer, and periodically draws/vkQueuePresentKHRs.
Obviously there must be some sort of synchronization mechanism to prevent the "graphics" thread from reading from the VkBuffer whilst the "state" thread is writing to it.
The only idea I have is to employ the usage of a host mutex fromvkQueueSubmit to vkWaitForFences in both threads.
I want to know, is there perhaps some other method that is more efficient or is this considered to be OK?
Try using semaphores. They are used to synchronize operations solely on the GPU, which is much more optimal than waiting in the app and submitting work after previous work is fully processed.
When You submit work You can provide a semaphore which gets signaled when this work is finished. When You submit another work You can provide the same semaphore on which the second batch should wait. Processing of the second batch will start automatically when the semaphore gets signaled (this semaphore is also automatically unsignaled and can be reused).
(I think there are some constraints on using semaphores, associated with queues. I will update the answer later when I confirm this but they should be sufficient for Your purposes.
[EDIT] There are constraints on using semaphores but it shouldn't affect You - when You use a semaphore as a wait semaphore during submission, no other queue can wait on the same semaphore.)
There are also events in Vulkan which can be used for similar purposes but their use is a little bit more complicated.
If You really need to synchronize GPU and Your application, use fences. They are signaled in a similar way as semaphores. But You can check their state on the app side and You need to manually unsignal them before You can use then again.
[EDIT]
I've added an image that more or less shows what I think You should do. One thread calculates state and with each submission adds a semaphore to the top of the list (or a ring buffer as #NicolasBolas wrote). This semaphore gets signaled when the submission is finished (it is provided in pSignalSemaphores during "compute" batch submission).
Second thread renders Your scene. It manages it's own list of semaphores similarly to the compute thread. But when You want to render things, You need to be sure that compute thread finished calculations. That's why You need to take the latest "compute" semaphore and wait on it (provide it in pWaitSemaphores during "render" batch submission). When You submit rendering commands, compute thread can't start and modify the data because it may influence the results of a rendering. So compute thread also needs to wait until the most recent rendering is done. That's why compute thread also needs to provide a wait semaphore (the most recent "rendering" semaphore).
You just need to synchronize submissions. Rendering thread cannot start when a compute threads submits commands and vice versa. That's why adding semaphores to the lists (and taking semaphores from the list) should be synchronized. But this has nothing to do with Vulkan. Probably some mutex will be helpful (for example a C++-ish std::lock_guard<std::mutex>). But this synchronization is a problem only when You have a single buffer.
Another thing is what to do with old semaphores from both lists. You cannot directly check what is their state and You cannot directly unsignal them. The state of semaphores can be checked by using additional fences provided with each submission. You don't wait on them but from time to time check if a given fence is signaled and, if it is, You can destroy old semaphore (as You cannot unsignal it from the application) or You can make an empty submission, with no command buffers, and use that semaphore as a wait semaphore. This way the semaphore will be unsignaled and You can reuse it. But I don't know which solution is more optimal: destroying old and creating new semaphores, or unsignaling them with empty submissions.
When You have a single buffer, a one-element list/ring is probably enough. But more optimal solution would have some kind of a ping-pong set of buffers - You read data from one buffer, but store results in another buffer. And in the next step You swap them. That's why in the image above, the lists of semaphores (rings) may have more elements depending on Your setup. The more independent buffers and semaphores in the lists (of course to some reasonable count), the best performance You will get as You reduce time wasted on waiting. But this complicates Your code and it may also increase a lag (rendering thread gets data that is a bit older than the data currently processed by the compute thread). So You may need to balance performance, code complexity and a rendering lag.
How you do this depends on two factors:
Whether you want to dispatch the compute operation on the same queue as its corresponding graphics operation.
The ratio of compute operations to their corresponding graphics operations.
#2 is the most important part.
Even though they are generated in separate threads, there must be at least some idea that the graphics operation is being fed by a particular compute operation (otherwise, how would the graphics thread know where the data is to read from?). So, how do you do that?
At the end of the day, that part has nothing to do with Vulkan. You need to use some inter-thread communication mechanism to allow the graphics thread to ask, "which compute task's data should I be using?"
Typically, this would be done by having the compute thread add every compute operation it does to some kind of circular buffer (thread-safe of course. And non-locking). When the graphics thread goes to decide where to read its data from, it asks the circular buffer for the most recently added compute operation.
In addition to the "where to read its data from" information, this would also provide the graphics thread with an appropriate Vulkan synchronization primitive to use to synchronize its command buffer(s) with the compute operation's CB.
If the compute and graphics operations are being dispatched on the same queue, then this is pretty simple. There doesn't have to actually be a synchronization primitive. So long as the graphics CBs are issued after the compute CBs in the batch, all the graphics CBs need is to have a vkCmdPipelineBarrier at the front which waits on all memory operations from the compute stage.
srcStageMask would be STAGE_COMPUTE_SHADER_BIT, with dstStageMask being, well, pretty much everything (you could narrow it down, but it won't matter, since at the very least your vertex shader stage will need to be there).
You would need a single VkMemoryBarrier in the pipeline barrier. It's srcAccessMask would be SHADER_WRITE_BIT, while the dstAccessMask would be however you intend to read it. If the compute operations wrote some vertex data, you need VERTEX_ATTRIBUTE_READ_BIT. If they wrote some uniform buffer data, you need UNIFORM_READ_BIT. And so on.
If you're dispatching these operations on separate queues, that's where you need an actual synchronization object.
There are several problems:
You cannot detect if a Vulkan semaphore has been signaled by user code. Nor can you set a semaphore to the unsignaled state by user code. Nor can you reasonably submit a batch that has a semaphore in it that is currently signaled and nobody's waiting on it. You can do the latter, but it won't do the right thing.
In short, you can never submit a batch that signals a semaphore unless you are certain that some process is going to wait for it.
You cannot issue a batch that waits on a semaphore, unless a batch that signals it is "pending execution". That is, your graphics thread cannot vkQueueSubmit its batch until it is certain that the compute queue has submitted its signaling batch.
So what you have to do is this. When the graphics queue goes to get its compute data, this must send a signal to the compute thread to add a semaphore to its next submit call. When the graphics thread submits its graphics operation, it then waits on that semaphore.
But to ensure proper ordering, the graphics thread cannot submit its operation until the compute thread has submitted the semaphore signaling operation. That requires a CPU-synchronization operation of some form. It could be as simple as the graphics thread polling an atomic variable set by the compute thread.
From this post:
Two threads being timesliced on a single CPU core won't run into a reordering problem. A single core always knows about its own reordering and will properly resolve all its own memory accesses. Multiple cores however operate independently in this regard and thus won't really know about each other's reordering.
Why can't the instruction reorder issue occur on a single CPU core? This article doesn't explain it.
EXAMPLE:
The following pictures are picked from Memory Reordering Caught in the Act:
Below is recorded:
I think the recorded instructions can also cause issue on a single CPU, because both r1 and r2 aren't 1.
A single core always knows about its own reordering and will properly resolve all its own memory accesses.
A single CPU core does reorder, but it knows it's own reordering, and can do clever tricks to pretend it's not. Thus, things go faster, without weird side effects.
Multiple cores however operate independently in this regard and thus won't really know about each other's reordering.
When a CPU reorders, the other CPUs can't compensate for this. Imagine if CPU #1 is waiting for a write to variableA, then it reads from variableB. If CPU#2 wrotes to variableB, then variableA like the code says, no problems occur. If CPU#2 reorders to write to variableA first, then CPU#1 doesn't know and tries to read from variableB before it has a value. This can cause crashes or any "random" behavior. (Intel chips have more magic that makes this not happen)
Two threads being timesliced on a single CPU core won't run into a reordering problem.
If both threads are on the same CPU, then it doesn't matter which order the writes happen in, because if they're reordered, then they're both in progress, and the CPU won't really switch until both are written, in which case they're safe to read from the other thread.
Example
For the code to have a problem on a single core, it would have to rearrange the two instructions from process 1 and be interrupted by process 2 and execute that between the two instructions. But if interrupted between them, it knows it has to abort both of them since it knows about it's own reordering, and knows it's in a dangerous state. So it will either do them in order, or do both before switching to process 2, or do neither before switching to process 2. All of which avoid the reordering problem.
There are multiple effects at work, but they are modeled as just one effect. Makes it easier to reason about them. Yes, a modern core already re-orders instructions by itself. But it maintains logical flow between them, if two instructions have an inter-dependency between them then they stay ordered so the logic of the program does not change. Discovering these inter-dependencies and preventing an instruction from being issued too early is the job of the reorder buffer in the execution engine.
This logic is solid and can be relied upon, it would be next to impossible to write a program if that wasn't the case. But that same guarantee cannot be provided by the memory controller. It has the un-enviable job of giving multiple processors access to the same shared memory.
First is the prefetcher, it reads data from memory ahead of time to ensure the data is available by the time a read instruction executes. Ensures the core won't stall waiting for the read to complete. With the problem that, because memory was read early, it might be a stale value that was changed by another core between the time the prefetch was done and the read instruction executes. To an outside observer it looks like the instruction executed early.
And the store buffer, it takes the data of a write instruction and writes it lazily to memory. Later, after the instruction executed. Ensures the core won't stall waiting on the memory bus write cycle to complete. To an outside observer, it just looks like the instruction executed late.
Modeling the effects of the prefetcher and store buffer as instruction reordering effects is very convenient. You can write that down on a piece of paper easily and reason about the side-effects.
To the core itself, the effects of the prefetcher and store buffer are entirely benign and it is oblivious to them. As long as there isn't another core that's also changing memory content. A machine with a single core always has that guarantee.
I've been reading about semaphores and came across this article:
www.csc.villanova.edu/~mdamian/threads/posixsem.html
So, this page states that if there are two threads accessing the same data, things can get ugly. The solution is to allow only one thread to access the data at the same time.
This is clear and I understand the solution, only why would anyone need threads to do this? What is the point? If the threads are blocked so that only one can execute, why use them at all? There is no advantage. (or maybe this is a just a dumb example; in such a case please point me to a sensible one)
Thanks in advance.
Consider this:
void update_shared_variable() {
sem_wait( &g_shared_variable_mutex );
g_shared_variable++;
sem_post( &g_shared_variable_mutex );
}
void thread1() {
do_thing_1a();
do_thing_1b();
do_thing_1c();
update_shared_variable(); // may block
}
void thread2() {
do_thing_2a();
do_thing_2b();
do_thing_2c();
update_shared_variable(); // may block
}
Note that all of the do_thing_xx functions still happen simultaneously. The semaphore only comes into play when the threads need to modify some shared (global) state or use some shared resource. So a thread will only block if another thread is trying to access the shared thing at the same time.
Now, if the only thing your threads are doing is working with one single shared variable/resource, then you are correct - there is no point in having threads at all (it would actually be less efficient than just one thread, due to context switching.)
When you are using multithreading not everycode that runs will be blocking. For example, if you had a queue, and two threads are reading from that queue, you would make sure that no thread reads at the same time from the queue, so that part would be blocking, but that's the part that will probably take the less time. Once you have retrieved the item to process from the queue, all the rest of the code can be run asynchronously.
The idea behind the threads is to allow simultaneous processing. A shared resource must be governed to avoid things like deadlocks or starvation. If something can take a while to process, then why not create multiple instances of those processes to allow them to finish faster? The bottleneck is just what you mentioned, when a process has to wait for I/O.
Being blocked while waiting for the shared resource is small when compared to the processing time, this is when you want to use multiple threads.
This is of course a SSCCE (Short, Self Contained, Correct Example)
Let's say you have 2 worker threads that do a lot of work and write the result to a file.
you only need to lock the file (shared resource) access.
The problem with trivial examples....
If the problem you're trying to solve can be broken down into pieces that can be executed in parallel then threads are a good thing.
A slightly less trivial example - imagine a for loop where the data being processed in each iteration is different every time. In that circumstance you could execute each iteration of the for loop simultaneously in separate threads. And indeed some compilers like Intel's will convert suitable for loops to threads automatically for you. In that particular circumstances no semaphores are needed because of the iterations' data independence.
But say you were wanting to process a stream of data, and that processing had two distinct steps, A and B. The threadless approach would involve reading in some data then doing A then B and then output the data before reading more input. Or you could have a thread reading and doing A, another thread doing B and output. So how do you get the interim result from the first thread to the second?
One way would be to have a memory buffer to contain the interim result. The first thread could write the interim result to a memory buffer and the second could read from it. But with two threads operating independently there's no way for the first thread to know if it's safe to overwrite that buffer, and there's no way for the second to know when to read from it.
That's where you can use semaphores to synchronise the action of the two threads. The first thread takes a semaphore that I'll call empty, fills the buffer, and then posts a semaphore called filled. Meanwhile the second thread will take the filled semaphore, read the buffer, and then post empty. So long as filled is initialised to 0 and empty is initialised to 1 it will work. The second thread will process the data only after the first has written it, and the first won't write it until the second has finished with it.
It's only worth it of course if the amount of time each thread spends processing data outweighs the amount of time spent waiting for semaphores. This limits the extent to which splitting code up into threads yields a benefit. Going beyond that tends to mean that the overall execution is effectively serial.
You can do multithreaded programming without semaphores at all. There's the Actor model or Communicating Sequential Processes (the one I favour). It's well worth looking up JCSP on Wikipedia.
In these programming styles data is shared between threads by sending it down communication channels. So instead of using semaphores to grant another thread access to data it would be sent a copy of that data down something a bit like a network socket, or a pipe. The advantage of CSP (which limits that communication channel to send-finishes-only-if-receiver-has-read) is that it stops you falling into the many many pitfalls that plague multithreaded do programs. It sounds inefficient (copying data is inefficient), but actually it's not so bad with Intel's QPI architecture, AMD's Hypertransport. And it means hat the 'channel' really could be a network connection; scalability built in by design.
I'm looking for real world examples of needing read and write access to the same value in concurrent systems.
In my opinion, many semaphores or locks are present because there's no known alternative (to the implementer,) but do you know of any patterns where mutexes seem to be a requirement?
In a way I'm asking for candidates for the standard set of HARD problems for concurrent software in the real world.
What kind of locks are used depends on how the data is being accessed by multiple threads. If you can fine tune the use case, you can sometimes eliminate the need for exclusive locks completely.
An exclusive lock is needed only if your use case requires that the shared data must be 100% exact all the time. This is the default that most developers start with because that's how we think about data normally.
However, if what you are using the data for can tolerate some "looseness", there are several techniques to share data between threads without the use of exclusive locks on every access.
For example, if you have a linked list of data and if your use of that linked list would not be upset by seeing the same node multiple times in a list traversal and would not be upset if it did not see an insert immediately after the insert (or similar artifacts), you can perform list inserts and deletes using atomic pointer exchange without the need for a full-stop mutex lock around the insert or delete operation.
Another example: if you have an array or list object that is mostly read from by threads and only occasionally updated by a master thread, you could implement lock-free updates by maintaining two copies of the list: one that is "live" that other threads can read from and another that is "offline" that you can write to in the privacy of your own thread. To perform an update, you copy the contents of the "live" list into the "offline" list, perform the update to the offline list, and then swap the offline list pointer into the live list pointer using an atomic pointer exchange. You will then need some mechanism to let the readers "drain" from the now offline list. In a garbage collected system, you can just release the reference to the offline list - when the last consumer is finished with it, it will be GC'd. In a non-GC system, you could use reference counting to keep track of how many readers are still using the list. For this example, having only one thread designated as the list updater would be ideal. If multiple updaters are needed, you will need to put a lock around the update operation, but only to serialize updaters - no lock and no performance impact on readers of the list.
All the lock-free resource sharing techniques I'm aware of require the use of atomic swaps (aka InterlockedExchange). This usually translates into a specific instruction in the CPU and/or a hardware bus lock (lock prefix on a read or write opcode in x86 assembler) for a very brief period of time. On multiproc systems, atomic swaps may force a cache invalidation on the other processors (this was the case on dual proc Pentium II) but I don't think this is as much of a problem on current multicore chips. Even with these performance caveats, lock-free runs much faster than taking a full-stop kernel event object. Just making a call into a kernel API function takes several hundred clock cycles (to switch to kernel mode).
Examples of real-world scenarios:
producer/consumer workflows. Web service receives http requests for data, places the request into an internal queue, worker thread pulls the work item from the queue and performs the work. The queue is read/write and has to be thread safe.
Data shared between threads with change of ownership. Thread 1 allocates an object, tosses it to thread 2 for processing, and never wants to see it again. Thread 2 is responsible for disposing the object. The memory management system (malloc/free) must be thread safe.
File system. This is almost always an OS service and already fully thread safe, but it's worth including in the list.
Reference counting. Releases the resource when the number of references drops to zero. The increment/decrement/test operations must be thread safe. These can usually be implemented using atomic primitives instead of full-stop kernal mutex locks.
Most real world, concurrent software, has some form of requirement for synchronization at some level. Often, better written software will take great pains to reduce the amount of locking required, but it is still required at some point.
For example, I often do simulations where we have some form of aggregation operation occurring. Typically, there are ways to prevent locking during the simulation phase itself (ie: use of thread local state data, etc), but the actual aggregation portion typically requires some form of lock at the end.
Luckily, this becomes a lock per thread, not per unit of work. In my case, this is significant, since I'm typically doing operations on hundreds of thousands or millions of units of work, but most of the time, it's occuring on systems with 4-16 PEs, which means I'm usually restricting to a similar number of units of execution. By using this type of mechanism, you're still locking, but you're locking between tens of elements instead of potentially millions.