Re: [PATCH 3/5] mm/vmalloc.c: correct lazy_max_pages() return value
On Fri, 23 Sep 2016 13:00:35 +0800 zijun_hu wrote: > On 2016/9/23 11:30, Nicholas Piggin wrote: > > On Fri, 23 Sep 2016 00:30:20 +0800 > > zijun_hu wrote: > > > >> On 2016/9/22 20:37, Michal Hocko wrote: > >>> On Thu 22-09-16 09:13:50, zijun_hu wrote: > On 09/22/2016 08:35 AM, David Rientjes wrote: > >>> [...] > > The intent is as it is implemented; with your change, lazy_max_pages() > > is > > potentially increased depending on the number of online cpus. This is > > only a heuristic, changing it would need justification on why the new > > value is better. It is opposite to what the comment says: "to be > > conservative and not introduce a big latency on huge systems, so go with > > a less aggressive log scale." NACK to the patch. > > > my change potentially make lazy_max_pages() decreased not increased, i > seems > conform with the comment > > if the number of online CPUs is not power of 2, both have no any > difference > otherwise, my change remain power of 2 value, and the original code > rounds up > to next power of 2 value, for instance > > my change : (32, 64] -> 64 > 32 -> 32, 64 -> 64 > the original code: [32, 63) -> 64 > 32 -> 64, 64 -> 128 > >>> > >>> You still completely failed to explain _why_ this is an improvement/fix > >>> or why it matters. This all should be in the changelog. > >>> > >> > >> Hi npiggin, > >> could you give some comments for this patch since lazy_max_pages() is > >> introduced > >> by you > >> > >> my patch is based on the difference between fls() and get_count_order() > >> mainly > >> the difference between fls() and get_count_order() will be shown below > >> more MM experts maybe help to decide which is more suitable > >> > >> if parameter > 1, both have different return value only when parameter is > >> power of two, for example > >> > >> fls(32) = 6 VS get_count_order(32) = 5 > >> fls(33) = 6 VS get_count_order(33) = 6 > >> fls(63) = 6 VS get_count_order(63) = 6 > >> fls(64) = 7 VS get_count_order(64) = 6 > >> > >> @@ -594,7 +594,9 @@ static unsigned long lazy_max_pages(void) > >> { > >> unsigned int log; > >> > >> -log = fls(num_online_cpus()); > >> +log = num_online_cpus(); > >> +if (log > 1) > >> +log = (unsigned int)get_count_order(log); > >> > >> return log * (32UL * 1024 * 1024 / PAGE_SIZE); > >> } > >> > > > > To be honest, I don't think I chose it with a lot of analysis. > > It will depend on the kernel usage patterns, the arch code, > > and the CPU microarchitecture, all of which would have changed > > significantly. > > > > I wouldn't bother changing it unless you do some bench marking > > on different system sizes to see where the best performance is. > > (If performance is equal, fewer lazy pages would be better.) > > > > Good to see you taking a look at this vmalloc stuff. Don't be > > discouraged if you run into some dead ends. > > > > Thanks, > > Nick > > > thanks for your reply > please don't pay attention to this patch any more since i don't have > condition to do many test and comparison > > i just feel my change maybe be consistent with operation of rounding up > to power of 2 You could be right about that, but in cases like this, existing code probably trumps original comments or intention. What I mean is that it's a heuristic anyway, so the current behaviour is what has been used and tested for a long time. Therefore any change would need to be justified by showing it's an improvement. I'm sure the existing size is not perfect, but we don't know whether your change would be an improvement in practice. Does that make sense? Thanks, Nick
Re: [PATCH 3/5] mm/vmalloc.c: correct lazy_max_pages() return value
On 2016/9/23 11:30, Nicholas Piggin wrote: > On Fri, 23 Sep 2016 00:30:20 +0800 > zijun_hu wrote: > >> On 2016/9/22 20:37, Michal Hocko wrote: >>> On Thu 22-09-16 09:13:50, zijun_hu wrote: On 09/22/2016 08:35 AM, David Rientjes wrote: >>> [...] > The intent is as it is implemented; with your change, lazy_max_pages() is > potentially increased depending on the number of online cpus. This is > only a heuristic, changing it would need justification on why the new > value is better. It is opposite to what the comment says: "to be > conservative and not introduce a big latency on huge systems, so go with > a less aggressive log scale." NACK to the patch. > my change potentially make lazy_max_pages() decreased not increased, i seems conform with the comment if the number of online CPUs is not power of 2, both have no any difference otherwise, my change remain power of 2 value, and the original code rounds up to next power of 2 value, for instance my change : (32, 64] -> 64 32 -> 32, 64 -> 64 the original code: [32, 63) -> 64 32 -> 64, 64 -> 128 >>> >>> You still completely failed to explain _why_ this is an improvement/fix >>> or why it matters. This all should be in the changelog. >>> >> >> Hi npiggin, >> could you give some comments for this patch since lazy_max_pages() is >> introduced >> by you >> >> my patch is based on the difference between fls() and get_count_order() >> mainly >> the difference between fls() and get_count_order() will be shown below >> more MM experts maybe help to decide which is more suitable >> >> if parameter > 1, both have different return value only when parameter is >> power of two, for example >> >> fls(32) = 6 VS get_count_order(32) = 5 >> fls(33) = 6 VS get_count_order(33) = 6 >> fls(63) = 6 VS get_count_order(63) = 6 >> fls(64) = 7 VS get_count_order(64) = 6 >> >> @@ -594,7 +594,9 @@ static unsigned long lazy_max_pages(void) >> { >> unsigned int log; >> >> -log = fls(num_online_cpus()); >> +log = num_online_cpus(); >> +if (log > 1) >> +log = (unsigned int)get_count_order(log); >> >> return log * (32UL * 1024 * 1024 / PAGE_SIZE); >> } >> > > To be honest, I don't think I chose it with a lot of analysis. > It will depend on the kernel usage patterns, the arch code, > and the CPU microarchitecture, all of which would have changed > significantly. > > I wouldn't bother changing it unless you do some bench marking > on different system sizes to see where the best performance is. > (If performance is equal, fewer lazy pages would be better.) > > Good to see you taking a look at this vmalloc stuff. Don't be > discouraged if you run into some dead ends. > > Thanks, > Nick > thanks for your reply please don't pay attention to this patch any more since i don't have condition to do many test and comparison i just feel my change maybe be consistent with operation of rounding up to power of 2
Re: [PATCH 3/5] mm/vmalloc.c: correct lazy_max_pages() return value
On Fri, 23 Sep 2016 00:30:20 +0800 zijun_hu wrote: > On 2016/9/22 20:37, Michal Hocko wrote: > > On Thu 22-09-16 09:13:50, zijun_hu wrote: > >> On 09/22/2016 08:35 AM, David Rientjes wrote: > > [...] > >>> The intent is as it is implemented; with your change, lazy_max_pages() is > >>> potentially increased depending on the number of online cpus. This is > >>> only a heuristic, changing it would need justification on why the new > >>> value is better. It is opposite to what the comment says: "to be > >>> conservative and not introduce a big latency on huge systems, so go with > >>> a less aggressive log scale." NACK to the patch. > >>> > >> my change potentially make lazy_max_pages() decreased not increased, i > >> seems > >> conform with the comment > >> > >> if the number of online CPUs is not power of 2, both have no any difference > >> otherwise, my change remain power of 2 value, and the original code rounds > >> up > >> to next power of 2 value, for instance > >> > >> my change : (32, 64] -> 64 > >> 32 -> 32, 64 -> 64 > >> the original code: [32, 63) -> 64 > >>32 -> 64, 64 -> 128 > > > > You still completely failed to explain _why_ this is an improvement/fix > > or why it matters. This all should be in the changelog. > > > > Hi npiggin, > could you give some comments for this patch since lazy_max_pages() is > introduced > by you > > my patch is based on the difference between fls() and get_count_order() mainly > the difference between fls() and get_count_order() will be shown below > more MM experts maybe help to decide which is more suitable > > if parameter > 1, both have different return value only when parameter is > power of two, for example > > fls(32) = 6 VS get_count_order(32) = 5 > fls(33) = 6 VS get_count_order(33) = 6 > fls(63) = 6 VS get_count_order(63) = 6 > fls(64) = 7 VS get_count_order(64) = 6 > > @@ -594,7 +594,9 @@ static unsigned long lazy_max_pages(void) > { > unsigned int log; > > -log = fls(num_online_cpus()); > +log = num_online_cpus(); > +if (log > 1) > +log = (unsigned int)get_count_order(log); > > return log * (32UL * 1024 * 1024 / PAGE_SIZE); > } > To be honest, I don't think I chose it with a lot of analysis. It will depend on the kernel usage patterns, the arch code, and the CPU microarchitecture, all of which would have changed significantly. I wouldn't bother changing it unless you do some benchmarking on different system sizes to see where the best performance is. (If performance is equal, fewer lazy pages would be better.) Good to see you taking a look at this vmalloc stuff. Don't be discouraged if you run into some dead ends. Thanks, Nick
Re: [PATCH 3/5] mm/vmalloc.c: correct lazy_max_pages() return value
On 2016/9/22 20:37, Michal Hocko wrote: > On Thu 22-09-16 09:13:50, zijun_hu wrote: >> On 09/22/2016 08:35 AM, David Rientjes wrote: > [...] >>> The intent is as it is implemented; with your change, lazy_max_pages() is >>> potentially increased depending on the number of online cpus. This is >>> only a heuristic, changing it would need justification on why the new >>> value is better. It is opposite to what the comment says: "to be >>> conservative and not introduce a big latency on huge systems, so go with >>> a less aggressive log scale." NACK to the patch. >>> >> my change potentially make lazy_max_pages() decreased not increased, i seems >> conform with the comment >> >> if the number of online CPUs is not power of 2, both have no any difference >> otherwise, my change remain power of 2 value, and the original code rounds up >> to next power of 2 value, for instance >> >> my change : (32, 64] -> 64 >> 32 -> 32, 64 -> 64 >> the original code: [32, 63) -> 64 >>32 -> 64, 64 -> 128 > > You still completely failed to explain _why_ this is an improvement/fix > or why it matters. This all should be in the changelog. > Hi npiggin, could you give some comments for this patch since lazy_max_pages() is introduced by you my patch is based on the difference between fls() and get_count_order() mainly the difference between fls() and get_count_order() will be shown below more MM experts maybe help to decide which is more suitable if parameter > 1, both have different return value only when parameter is power of two, for example fls(32) = 6 VS get_count_order(32) = 5 fls(33) = 6 VS get_count_order(33) = 6 fls(63) = 6 VS get_count_order(63) = 6 fls(64) = 7 VS get_count_order(64) = 6 @@ -594,7 +594,9 @@ static unsigned long lazy_max_pages(void) { unsigned int log; -log = fls(num_online_cpus()); +log = num_online_cpus(); +if (log > 1) +log = (unsigned int)get_count_order(log); return log * (32UL * 1024 * 1024 / PAGE_SIZE); }
Re: [PATCH 3/5] mm/vmalloc.c: correct lazy_max_pages() return value
On Thu 22-09-16 09:13:50, zijun_hu wrote: > On 09/22/2016 08:35 AM, David Rientjes wrote: [...] > > The intent is as it is implemented; with your change, lazy_max_pages() is > > potentially increased depending on the number of online cpus. This is > > only a heuristic, changing it would need justification on why the new > > value is better. It is opposite to what the comment says: "to be > > conservative and not introduce a big latency on huge systems, so go with > > a less aggressive log scale." NACK to the patch. > > > my change potentially make lazy_max_pages() decreased not increased, i seems > conform with the comment > > if the number of online CPUs is not power of 2, both have no any difference > otherwise, my change remain power of 2 value, and the original code rounds up > to next power of 2 value, for instance > > my change : (32, 64] -> 64 >32 -> 32, 64 -> 64 > the original code: [32, 63) -> 64 >32 -> 64, 64 -> 128 You still completely failed to explain _why_ this is an improvement/fix or why it matters. This all should be in the changelog. -- Michal Hocko SUSE Labs
Re: [PATCH 3/5] mm/vmalloc.c: correct lazy_max_pages() return value
On 09/22/2016 08:35 AM, David Rientjes wrote: > On Thu, 22 Sep 2016, zijun_hu wrote: > >> On 2016/9/22 5:21, David Rientjes wrote: >>> On Wed, 21 Sep 2016, zijun_hu wrote: >>> From: zijun_hu correct lazy_max_pages() return value if the number of online CPUs is power of 2 Signed-off-by: zijun_hu --- mm/vmalloc.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/mm/vmalloc.c b/mm/vmalloc.c index a125ae8..2804224 100644 --- a/mm/vmalloc.c +++ b/mm/vmalloc.c @@ -594,7 +594,9 @@ static unsigned long lazy_max_pages(void) { unsigned int log; - log = fls(num_online_cpus()); + log = num_online_cpus(); + if (log > 1) + log = (unsigned int)get_count_order(log); return log * (32UL * 1024 * 1024 / PAGE_SIZE); } >>> >>> The implementation of lazy_max_pages() is somewhat arbitrarily defined, >>> the existing approximation has been around for eight years and >>> num_online_cpus() isn't intended to be rounded up to the next power of 2. >>> I'd be inclined to just leave it as it is. >>> >> do i understand the intent in current code logic as below ? >> [8, 15) roundup to 16? >> [32, 63) roundup to 64? >> > > The intent is as it is implemented; with your change, lazy_max_pages() is > potentially increased depending on the number of online cpus. This is > only a heuristic, changing it would need justification on why the new > value is better. It is opposite to what the comment says: "to be > conservative and not introduce a big latency on huge systems, so go with > a less aggressive log scale." NACK to the patch. > my change potentially make lazy_max_pages() decreased not increased, i seems conform with the comment if the number of online CPUs is not power of 2, both have no any difference otherwise, my change remain power of 2 value, and the original code rounds up to next power of 2 value, for instance my change : (32, 64] -> 64 32 -> 32, 64 -> 64 the original code: [32, 63) -> 64 32 -> 64, 64 -> 128
Re: [PATCH 3/5] mm/vmalloc.c: correct lazy_max_pages() return value
On Thu, 22 Sep 2016, zijun_hu wrote: > On 2016/9/22 5:21, David Rientjes wrote: > > On Wed, 21 Sep 2016, zijun_hu wrote: > > > >> From: zijun_hu > >> > >> correct lazy_max_pages() return value if the number of online > >> CPUs is power of 2 > >> > >> Signed-off-by: zijun_hu > >> --- > >> mm/vmalloc.c | 4 +++- > >> 1 file changed, 3 insertions(+), 1 deletion(-) > >> > >> diff --git a/mm/vmalloc.c b/mm/vmalloc.c > >> index a125ae8..2804224 100644 > >> --- a/mm/vmalloc.c > >> +++ b/mm/vmalloc.c > >> @@ -594,7 +594,9 @@ static unsigned long lazy_max_pages(void) > >> { > >>unsigned int log; > >> > >> - log = fls(num_online_cpus()); > >> + log = num_online_cpus(); > >> + if (log > 1) > >> + log = (unsigned int)get_count_order(log); > >> > >>return log * (32UL * 1024 * 1024 / PAGE_SIZE); > >> } > > > > The implementation of lazy_max_pages() is somewhat arbitrarily defined, > > the existing approximation has been around for eight years and > > num_online_cpus() isn't intended to be rounded up to the next power of 2. > > I'd be inclined to just leave it as it is. > > > do i understand the intent in current code logic as below ? > [8, 15) roundup to 16? > [32, 63) roundup to 64? > The intent is as it is implemented; with your change, lazy_max_pages() is potentially increased depending on the number of online cpus. This is only a heuristic, changing it would need justification on why the new value is better. It is opposite to what the comment says: "to be conservative and not introduce a big latency on huge systems, so go with a less aggressive log scale." NACK to the patch.
Re: [PATCH 3/5] mm/vmalloc.c: correct lazy_max_pages() return value
On 2016/9/22 5:21, David Rientjes wrote: > On Wed, 21 Sep 2016, zijun_hu wrote: > >> From: zijun_hu >> >> correct lazy_max_pages() return value if the number of online >> CPUs is power of 2 >> >> Signed-off-by: zijun_hu >> --- >> mm/vmalloc.c | 4 +++- >> 1 file changed, 3 insertions(+), 1 deletion(-) >> >> diff --git a/mm/vmalloc.c b/mm/vmalloc.c >> index a125ae8..2804224 100644 >> --- a/mm/vmalloc.c >> +++ b/mm/vmalloc.c >> @@ -594,7 +594,9 @@ static unsigned long lazy_max_pages(void) >> { >> unsigned int log; >> >> -log = fls(num_online_cpus()); >> +log = num_online_cpus(); >> +if (log > 1) >> +log = (unsigned int)get_count_order(log); >> >> return log * (32UL * 1024 * 1024 / PAGE_SIZE); >> } > > The implementation of lazy_max_pages() is somewhat arbitrarily defined, > the existing approximation has been around for eight years and > num_online_cpus() isn't intended to be rounded up to the next power of 2. > I'd be inclined to just leave it as it is. > do i understand the intent in current code logic as below ? [8, 15) roundup to 16? [32, 63) roundup to 64?
Re: [PATCH 3/5] mm/vmalloc.c: correct lazy_max_pages() return value
On Wed, 21 Sep 2016, zijun_hu wrote: > From: zijun_hu > > correct lazy_max_pages() return value if the number of online > CPUs is power of 2 > > Signed-off-by: zijun_hu > --- > mm/vmalloc.c | 4 +++- > 1 file changed, 3 insertions(+), 1 deletion(-) > > diff --git a/mm/vmalloc.c b/mm/vmalloc.c > index a125ae8..2804224 100644 > --- a/mm/vmalloc.c > +++ b/mm/vmalloc.c > @@ -594,7 +594,9 @@ static unsigned long lazy_max_pages(void) > { > unsigned int log; > > - log = fls(num_online_cpus()); > + log = num_online_cpus(); > + if (log > 1) > + log = (unsigned int)get_count_order(log); > > return log * (32UL * 1024 * 1024 / PAGE_SIZE); > } The implementation of lazy_max_pages() is somewhat arbitrarily defined, the existing approximation has been around for eight years and num_online_cpus() isn't intended to be rounded up to the next power of 2. I'd be inclined to just leave it as it is.