| Message ID | 20170828035327.17146-2-bobby.prani@gmail.com |
|---|---|
| State | New |
| Headers | show |
| Series | None | expand |
On 08/27/2017 08:53 PM, Pranith Kumar wrote: > Using heaptrack, I found that quite a few of our temporary allocations > are coming from allocating work items. Instead of doing this > continously, we can cache the allocated items and reuse them instead > of freeing them. > > This reduces the number of allocations by 25% (200000 -> 150000 for > ARM64 boot+shutdown test). > > Signed-off-by: Pranith Kumar <bobby.prani@gmail.com> > --- > cpus-common.c | 85 ++++++++++++++++++++++++++++++++++++++++++++++++----------- > 1 file changed, 70 insertions(+), 15 deletions(-) > > diff --git a/cpus-common.c b/cpus-common.c > index 59f751ecf9..a1c4c7d1a3 100644 > --- a/cpus-common.c > +++ b/cpus-common.c > @@ -24,6 +24,7 @@ > #include "sysemu/cpus.h" > > static QemuMutex qemu_cpu_list_lock; > +static QemuMutex qemu_wi_pool_lock; > static QemuCond exclusive_cond; > static QemuCond exclusive_resume; > static QemuCond qemu_work_cond; > @@ -33,6 +34,58 @@ static QemuCond qemu_work_cond; > */ > static int pending_cpus; > > +typedef struct qemu_work_item { > + struct qemu_work_item *next; > + run_on_cpu_func func; > + run_on_cpu_data data; > + bool free, exclusive, done; > +} qemu_work_item; > + > +typedef struct qemu_wi_pool { > + qemu_work_item *first, *last; > +} qemu_wi_pool; > + > +qemu_wi_pool *wi_free_pool; > + > +static void qemu_init_workitem_pool(void) > +{ > + wi_free_pool = g_malloc0(sizeof(qemu_wi_pool)); > + wi_free_pool->first = NULL; > + wi_free_pool->last = NULL; > +} > + > +static void qemu_wi_pool_insert(qemu_work_item *item) > +{ > + qemu_mutex_lock(&qemu_wi_pool_lock); > + if (wi_free_pool->last == NULL) { > + wi_free_pool->first = item; > + wi_free_pool->last = item; > + } else { > + wi_free_pool->last->next = item; > + wi_free_pool->last = item; > + } > + qemu_mutex_unlock(&qemu_wi_pool_lock); > +} > + > +static qemu_work_item* qemu_wi_pool_remove(void) > +{ > + qemu_mutex_lock(&qemu_wi_pool_lock); > + qemu_work_item *ret = wi_free_pool->first; > + > + if (ret == NULL) > + goto out; > + > + wi_free_pool->first = ret->next; > + if (wi_free_pool->last == ret) { > + wi_free_pool->last = NULL; > + } Why does this list need to record a "last" element? It would seem a simple lifo would be sufficient. (You would also be able to manage the list via cmpxchg without a separate lock, but perhaps the difference between the two isn't measurable.) r~
On Sun, Aug 27, 2017 at 23:53:25 -0400, Pranith Kumar wrote: > Using heaptrack, I found that quite a few of our temporary allocations > are coming from allocating work items. Instead of doing this > continously, we can cache the allocated items and reuse them instead > of freeing them. > > This reduces the number of allocations by 25% (200000 -> 150000 for > ARM64 boot+shutdown test). > But what is the perf difference, if any? Adding a lock (or a cmpxchg) here is not a great idea. However, this is not yet immediately obvious because of other scalability bottlenecks. (if you boot many arm64 cores you'll see most of the time is spent idling on the BQL, see https://lists.gnu.org/archive/html/qemu-devel/2017-08/msg05207.html ) You're most likely better off using glib's slices, see https://developer.gnome.org/glib/stable/glib-Memory-Slices.html These slices use per-thread lists, so scalability should be OK. I also suggest profiling with either or both of jemalloc/tcmalloc (build with --enable-jemalloc/tcmalloc) in addition to using glibc's allocator, and then based on perf numbers decide whether this is something worth optimizing. Thanks, Emilio
On Mon, Aug 28, 2017 at 1:47 PM, Richard Henderson <richard.henderson@linaro.org> wrote: > On 08/27/2017 08:53 PM, Pranith Kumar wrote: >> Using heaptrack, I found that quite a few of our temporary allocations >> are coming from allocating work items. Instead of doing this >> continously, we can cache the allocated items and reuse them instead >> of freeing them. >> >> This reduces the number of allocations by 25% (200000 -> 150000 for >> ARM64 boot+shutdown test). >> >> Signed-off-by: Pranith Kumar <bobby.prani@gmail.com> > > Why does this list need to record a "last" element? > It would seem a simple lifo would be sufficient. > > (You would also be able to manage the list via cmpxchg without a separate lock, > but perhaps the difference between the two isn't measurable.) > Yes, seems like a better design choice. Will fix in next iteration. Thanks,
On Mon, Aug 28, 2017 at 3:05 PM, Emilio G. Cota <cota@braap.org> wrote: > On Sun, Aug 27, 2017 at 23:53:25 -0400, Pranith Kumar wrote: >> Using heaptrack, I found that quite a few of our temporary allocations >> are coming from allocating work items. Instead of doing this >> continously, we can cache the allocated items and reuse them instead >> of freeing them. >> >> This reduces the number of allocations by 25% (200000 -> 150000 for >> ARM64 boot+shutdown test). >> > > But what is the perf difference, if any? > > Adding a lock (or a cmpxchg) here is not a great idea. However, this is not yet > immediately obvious because of other scalability bottlenecks. (if > you boot many arm64 cores you'll see most of the time is spent idling > on the BQL, see > https://lists.gnu.org/archive/html/qemu-devel/2017-08/msg05207.html ) > > You're most likely better off using glib's slices, see > https://developer.gnome.org/glib/stable/glib-Memory-Slices.html > These slices use per-thread lists, so scalability should be OK. I think we should modify our g_malloc() to internally use this. Seems like an idea worth trying out. > > I also suggest profiling with either or both of jemalloc/tcmalloc > (build with --enable-jemalloc/tcmalloc) in addition to using glibc's > allocator, and then based on perf numbers decide whether this is something > worth optimizing. > OK, I will try to get some perf numbers.
Il 28 ago 2017 11:43 PM, "Pranith Kumar" <bobby.prani@gmail.com> ha scritto: On Mon, Aug 28, 2017 at 1:47 PM, Richard Henderson <richard.henderson@linaro.org> wrote: > On 08/27/2017 08:53 PM, Pranith Kumar wrote: >> Using heaptrack, I found that quite a few of our temporary allocations >> are coming from allocating work items. Instead of doing this >> continously, we can cache the allocated items and reuse them instead >> of freeing them. >> >> This reduces the number of allocations by 25% (200000 -> 150000 for >> ARM64 boot+shutdown test). >> >> Signed-off-by: Pranith Kumar <bobby.prani@gmail.com> > > Why does this list need to record a "last" element? > It would seem a simple lifo would be sufficient. > > (You would also be able to manage the list via cmpxchg without a separate lock, > but perhaps the difference between the two isn't measurable.) > Yes, seems like a better design choice. Will fix in next iteration. More recent glibc will also have an efficient per-thread allocator, and though I haven't yet benchmarked the newer glibc malloc, GSlice is slower than at least both tcmalloc and jemalloc. Perhaps you could instead make work items statically allocated? Thanks, Paolo Thanks, -- Pranith
diff --git a/cpus-common.c b/cpus-common.c index 59f751ecf9..a1c4c7d1a3 100644 --- a/cpus-common.c +++ b/cpus-common.c @@ -24,6 +24,7 @@ #include "sysemu/cpus.h" static QemuMutex qemu_cpu_list_lock; +static QemuMutex qemu_wi_pool_lock; static QemuCond exclusive_cond; static QemuCond exclusive_resume; static QemuCond qemu_work_cond; @@ -33,6 +34,58 @@ static QemuCond qemu_work_cond; */ static int pending_cpus; +typedef struct qemu_work_item { + struct qemu_work_item *next; + run_on_cpu_func func; + run_on_cpu_data data; + bool free, exclusive, done; +} qemu_work_item; + +typedef struct qemu_wi_pool { + qemu_work_item *first, *last; +} qemu_wi_pool; + +qemu_wi_pool *wi_free_pool; + +static void qemu_init_workitem_pool(void) +{ + wi_free_pool = g_malloc0(sizeof(qemu_wi_pool)); + wi_free_pool->first = NULL; + wi_free_pool->last = NULL; +} + +static void qemu_wi_pool_insert(qemu_work_item *item) +{ + qemu_mutex_lock(&qemu_wi_pool_lock); + if (wi_free_pool->last == NULL) { + wi_free_pool->first = item; + wi_free_pool->last = item; + } else { + wi_free_pool->last->next = item; + wi_free_pool->last = item; + } + qemu_mutex_unlock(&qemu_wi_pool_lock); +} + +static qemu_work_item* qemu_wi_pool_remove(void) +{ + qemu_mutex_lock(&qemu_wi_pool_lock); + qemu_work_item *ret = wi_free_pool->first; + + if (ret == NULL) + goto out; + + wi_free_pool->first = ret->next; + if (wi_free_pool->last == ret) { + wi_free_pool->last = NULL; + } + ret->next = NULL; + + out: + qemu_mutex_unlock(&qemu_wi_pool_lock); + return ret; +} + void qemu_init_cpu_list(void) { /* This is needed because qemu_init_cpu_list is also called by the @@ -43,6 +96,9 @@ void qemu_init_cpu_list(void) qemu_cond_init(&exclusive_cond); qemu_cond_init(&exclusive_resume); qemu_cond_init(&qemu_work_cond); + + qemu_init_workitem_pool(); + qemu_mutex_init(&qemu_wi_pool_lock); } void cpu_list_lock(void) @@ -106,14 +162,7 @@ void cpu_list_remove(CPUState *cpu) qemu_mutex_unlock(&qemu_cpu_list_lock); } -struct qemu_work_item { - struct qemu_work_item *next; - run_on_cpu_func func; - run_on_cpu_data data; - bool free, exclusive, done; -}; - -static void queue_work_on_cpu(CPUState *cpu, struct qemu_work_item *wi) +static void queue_work_on_cpu(CPUState *cpu, qemu_work_item *wi) { qemu_mutex_lock(&cpu->work_mutex); if (cpu->queued_work_first == NULL) { @@ -132,7 +181,7 @@ static void queue_work_on_cpu(CPUState *cpu, struct qemu_work_item *wi) void do_run_on_cpu(CPUState *cpu, run_on_cpu_func func, run_on_cpu_data data, QemuMutex *mutex) { - struct qemu_work_item wi; + qemu_work_item wi; if (qemu_cpu_is_self(cpu)) { func(cpu, data); @@ -145,6 +194,7 @@ void do_run_on_cpu(CPUState *cpu, run_on_cpu_func func, run_on_cpu_data data, wi.free = false; wi.exclusive = false; + queue_work_on_cpu(cpu, &wi); while (!atomic_mb_read(&wi.done)) { CPUState *self_cpu = current_cpu; @@ -156,9 +206,11 @@ void do_run_on_cpu(CPUState *cpu, run_on_cpu_func func, run_on_cpu_data data, void async_run_on_cpu(CPUState *cpu, run_on_cpu_func func, run_on_cpu_data data) { - struct qemu_work_item *wi; + qemu_work_item *wi = qemu_wi_pool_remove(); - wi = g_malloc0(sizeof(struct qemu_work_item)); + if (!wi) { + wi = g_malloc0(sizeof(qemu_work_item)); + } wi->func = func; wi->data = data; wi->free = true; @@ -299,9 +351,11 @@ void cpu_exec_end(CPUState *cpu) void async_safe_run_on_cpu(CPUState *cpu, run_on_cpu_func func, run_on_cpu_data data) { - struct qemu_work_item *wi; + qemu_work_item *wi = qemu_wi_pool_remove(); - wi = g_malloc0(sizeof(struct qemu_work_item)); + if (!wi) { + wi = g_malloc0(sizeof(qemu_work_item)); + } wi->func = func; wi->data = data; wi->free = true; @@ -312,7 +366,7 @@ void async_safe_run_on_cpu(CPUState *cpu, run_on_cpu_func func, void process_queued_cpu_work(CPUState *cpu) { - struct qemu_work_item *wi; + qemu_work_item *wi; if (cpu->queued_work_first == NULL) { return; @@ -343,7 +397,8 @@ void process_queued_cpu_work(CPUState *cpu) } qemu_mutex_lock(&cpu->work_mutex); if (wi->free) { - g_free(wi); + memset(wi, 0, sizeof(qemu_work_item)); + qemu_wi_pool_insert(wi); } else { atomic_mb_set(&wi->done, true); }
Using heaptrack, I found that quite a few of our temporary allocations are coming from allocating work items. Instead of doing this continously, we can cache the allocated items and reuse them instead of freeing them. This reduces the number of allocations by 25% (200000 -> 150000 for ARM64 boot+shutdown test). Signed-off-by: Pranith Kumar <bobby.prani@gmail.com> --- cpus-common.c | 85 ++++++++++++++++++++++++++++++++++++++++++++++++----------- 1 file changed, 70 insertions(+), 15 deletions(-)