search for: contend

A major problem with the queue spinlock patch is its performance at low contention level (2-4 contending tasks) where it is slower than the corresponding ticket spinlock code. The following table shows the execution time (in ms) of a micro-benchmark where 5M iterations of the lock/unlock cycles were run on a 10-core Westere-EX x86-64 CPU with 2 different types loads - standalone (lock and protected...

A major problem with the queue spinlock patch is its performance at low contention level (2-4 contending tasks) where it is slower than the corresponding ticket spinlock code path. The following table shows the execution time (in ms) of a micro-benchmark where 5M iterations of the lock/unlock cycles were run on a 10-core Westere-EX CPU with 2 different types loads - standalone (lock and protected d...

A major problem with the queue spinlock patch is its performance at low contention level (2-4 contending tasks) where it is slower than the corresponding ticket spinlock code path. The following table shows the execution time (in ms) of a micro-benchmark where 5M iterations of the lock/unlock cycles were run on a 10-core Westere-EX CPU with 2 different types loads - standalone (lock and protected d...

...ed (much) so using > it won't be a problem and you get to have arbitrary atomic ops. > >> Given that speed at low contention level which is the common case is >> important to get this patch accepted, I have to do what I can to make it run >> as far as possible for this 2 contending task case. > What I'm saying is that you can do the whole thing with a single > cmpxchg. No extra ops needed. And at that point you don't need a whole > byte, you can use a single bit. > > that removes the whole NR_CPUS dependent logic. After modifying it to do a determin...

...39;t have to be retried (much) so using it won't be a problem and you get to have arbitrary atomic ops. > Given that speed at low contention level which is the common case is > important to get this patch accepted, I have to do what I can to make it run > as far as possible for this 2 contending task case. What I'm saying is that you can do the whole thing with a single cmpxchg. No extra ops needed. And at that point you don't need a whole byte, you can use a single bit. that removes the whole NR_CPUS dependent logic. > >>+ /* > >>+ * Someone has steal th...

...39;t have to be retried (much) so using it won't be a problem and you get to have arbitrary atomic ops. > Given that speed at low contention level which is the common case is > important to get this patch accepted, I have to do what I can to make it run > as far as possible for this 2 contending task case. What I'm saying is that you can do the whole thing with a single cmpxchg. No extra ops needed. And at that point you don't need a whole byte, you can use a single bit. that removes the whole NR_CPUS dependent logic. > >>+ /* > >>+ * Someone has steal th...

...ntion is also higher because of > slower inter-node memory traffic. > > Due to the fact that spinlocks are acquired with preemption disabled, > the process will not be migrated to another CPU while it is trying > to get a spinlock. Ignoring interrupt handling, a CPU can only be > contending in one spinlock at any one time. Counting soft IRQ, hard > IRQ and NMI, a CPU can only have a maximum of 4 concurrent lock waiting > activities. By allocating a set of per-cpu queue nodes and used them > to form a waiting queue, we can encode the queue node address into a > much sma...

...ntion is also higher because of > slower inter-node memory traffic. > > Due to the fact that spinlocks are acquired with preemption disabled, > the process will not be migrated to another CPU while it is trying > to get a spinlock. Ignoring interrupt handling, a CPU can only be > contending in one spinlock at any one time. Counting soft IRQ, hard > IRQ and NMI, a CPU can only have a maximum of 4 concurrent lock waiting > activities. By allocating a set of per-cpu queue nodes and used them > to form a waiting queue, we can encode the queue node address into a > much sma...

...s as possible. If I have to do an atomic cmpxchg, it probably won't be faster than the regular qspinlock slowpath. Given that speed at low contention level which is the common case is important to get this patch accepted, I have to do what I can to make it run as far as possible for this 2 contending task case. >> + return 1; >> + } else if (old == _QSPINLOCK_LOCKED) { >> +try_again: >> + /* >> + * Wait until the lock byte is cleared to get the lock >> + */ >> + do { >> + cpu_relax(); >> + } while (ACCESS_ONCE(qlock->lock));...

On Feb 28, 2014 1:30 AM, "Peter Zijlstra" <peterz at infradead.org> wrote: > > At low contention the cmpxchg won't have to be retried (much) so using > it won't be a problem and you get to have arbitrary atomic ops. Peter, the difference between an atomic op and *no* atomic op is huge. And Waiman posted numbers for the optimization. Why do you argue with

On 02/26, Waiman Long wrote: > > @@ -144,7 +317,7 @@ static __always_inline int queue_spin_setlock(struct qspinlock *lock) > int qlcode = atomic_read(lock->qlcode); > > if (!(qlcode & _QSPINLOCK_LOCKED) && (atomic_cmpxchg(&lock->qlcode, > - qlcode, qlcode|_QSPINLOCK_LOCKED) == qlcode)) > + qlcode, code|_QSPINLOCK_LOCKED) == qlcode)) Hmm.

Updated version, this includes numbers for my SNB desktop and Waiman's variant. Curiously Waiman's version seems consistently slower on 2 cross node CPUs. Whereas my version seems to have a problem on SNB with 2 CPUs. There's something weird with the ticket lock numbers; when I compile the code with: gcc (Debian 4.7.2-5) 4.7.2 I get the first set; when I compile with: gcc

Peter, I was trying to implement the generic queue code exchange code using cmpxchg as suggested by you. However, when I gathered the performance data, the code performed worse than I expected at a higher contention level. Below were the execution time of the benchmark tool that I sent you: [xchg] [cmpxchg] # of tasks Ticket lock Queue lock Queue Lock

On Tue, Mar 04, 2014 at 12:48:26PM -0500, Waiman Long wrote: > Peter, > > I was trying to implement the generic queue code exchange code using > cmpxchg as suggested by you. However, when I gathered the performance > data, the code performed worse than I expected at a higher contention > level. Below were the execution time of the benchmark tool that I sent > you: > >

On Wed, Mar 12, 2014 at 03:08:24PM -0400, Waiman Long wrote: > On 03/12/2014 02:54 PM, Waiman Long wrote: > >+ /* > >+ * Set the lock bit& clear the waiting bit simultaneously > >+ * It is assumed that there is no lock stealing with this > >+ * quick path active. > >+ * > >+ * A direct memory store of _QSPINLOCK_LOCKED into the > >+ *

On 02/26, Waiman Long wrote: > > @@ -144,7 +317,7 @@ static __always_inline int queue_spin_setlock(struct qspinlock *lock) > int qlcode = atomic_read(lock->qlcode); > > if (!(qlcode & _QSPINLOCK_LOCKED) && (atomic_cmpxchg(&lock->qlcode, > - qlcode, qlcode|_QSPINLOCK_LOCKED) == qlcode)) > + qlcode, code|_QSPINLOCK_LOCKED) == qlcode)) Hmm.

On Tue, Mar 04, 2014 at 12:48:26PM -0500, Waiman Long wrote: > Peter, > > I was trying to implement the generic queue code exchange code using > cmpxchg as suggested by you. However, when I gathered the performance > data, the code performed worse than I expected at a higher contention > level. Below were the execution time of the benchmark tool that I sent > you: > >

Hi All, I''m trying to create a contended section of bandwidth using IMQ. I have the imq0 device up and running, with traffic passing through it. Firstly, I need to throttle the entire device imq0 to 2mbit/s. I would then like to add throttle rules for individual IP addresses, allowing them to pass up to 512kbit/s each, as long as im...

...360.083940 2 - nodes: 2 - nodes: 2: 1509.193824 2: 1209.090219 4: 48154.495998 4: 48547.242379 8: 137946.787244 8: 141381.498125 --- There a few curious facts I found (assuming my test code is sane). - Intel seems to be an order of magnitude faster on uncontended LOCKed ops compared to AMD - On Intel the uncontended qspinlock fast path (cmpxchg) seems slower than the uncontended ticket xadd -- although both are plenty fast when compared to AMD. - In general, replacing cmpxchg loops with unconditional atomic ops doesn't seem to matter a...

...360.083940 2 - nodes: 2 - nodes: 2: 1509.193824 2: 1209.090219 4: 48154.495998 4: 48547.242379 8: 137946.787244 8: 141381.498125 --- There a few curious facts I found (assuming my test code is sane). - Intel seems to be an order of magnitude faster on uncontended LOCKed ops compared to AMD - On Intel the uncontended qspinlock fast path (cmpxchg) seems slower than the uncontended ticket xadd -- although both are plenty fast when compared to AMD. - In general, replacing cmpxchg loops with unconditional atomic ops doesn't seem to matter a...