| Message ID | 20191211181016.14366-3-sjpark@amazon.de (mailing list archive) |
|---|---|
| State | Superseded |
| Headers | show |
| Series | xenbus/backend: Add a memory pressure handler callback | expand |
Hello, Please make sure you Cc me in blkback related patches. On Wed, Dec 11, 2019 at 06:10:15PM +0000, SeongJae Park wrote: > Each `blkif` has a free pages pool for the grant mapping. The size of > the pool starts from zero and be increased on demand while processing ^ is > the I/O requests. If current I/O requests handling is finished or 100 > milliseconds has passed since last I/O requests handling, it checks and > shrinks the pool to not exceed the size limit, `max_buffer_pages`. > > Therefore, host administrators can cause memory pressure in blkback by > attaching a large number of block devices and inducing I/O. Such > problematic situations can be avoided by limiting the maximum number of > devices that can be attached, but finding the optimal limit is not so > easy. Improper set of the limit can results in the memory pressure or a ^ s/the// > resource underutilization. This commit avoids such problematic > situations by squeezing the pools (returns every free page in the pool > to the system) for a while (users can set this duration via a module > parameter) if a memory pressure is detected. ^ s/a// > > Discussions > =========== > > The `blkback`'s original shrinking mechanism returns only pages in the > pool, which are not currently be used by `blkback`, to the system. In I think you can remove both comas in the above sentence. > other words, the pages that are not mapped with granted pages. Because > this commit is changing only the shrink limit but still uses the same > freeing mechanism it does not touch pages which are currently mapping > grants. > > Once a memory pressure is detected, this commit keeps the squeezing ^ s/a// > limit for a user-specified time duration. The duration should be > neither too long nor too short. If it is too long, the squeezing > incurring overhead can reduce the I/O performance. If it is too short, > `blkback` will not free enough pages to reduce the memory pressure. > This commit sets the value as `10 milliseconds` by default because it is > a short time in terms of I/O while it is a long time in terms of memory > operations. Also, as the original shrinking mechanism works for at > least every 100 milliseconds, this could be a somewhat reasonable > choice. I also tested other durations (refer to the below section for > more details) and confirmed that 10 milliseconds is the one that works > best with the test. That said, the proper duration depends on actual > configurations and workloads. That's why this commit allows users to > set the duration as a module parameter. > > Memory Pressure Test > ==================== > > To show how this commit fixes the memory pressure situation well, I > configured a test environment on a xen-running virtualization system. > On the `blkfront` running guest instances, I attach a large number of > network-backed volume devices and induce I/O to those. Meanwhile, I > measure the number of pages that swapped in (pswpin) and out (pswpout) > on the `blkback` running guest. The test ran twice, once for the > `blkback` before this commit and once for that after this commit. As > shown below, this commit has dramatically reduced the memory pressure: > > pswpin pswpout > before 76,672 185,799 > after 212 3,325 > > Optimal Aggressive Shrinking Duration > ------------------------------------- > > To find a best squeezing duration, I repeated the test with three > different durations (1ms, 10ms, and 100ms). The results are as below: > > duration pswpin pswpout > 1 852 6,424 > 10 212 3,325 > 100 203 3,340 > > As expected, the memory pressure has decreased as the duration is > increased, but the reduction stopped from the `10ms`. Based on this > results, I chose the default duration as 10ms. > > Performance Overhead Test > ========================= > > This commit could incur I/O performance degradation under severe memory > pressure because the squeezing will require more page allocations per > I/O. To show the overhead, I artificially made a worst-case squeezing > situation and measured the I/O performance of a `blkfront` running > guest. > > For the artificial squeezing, I set the `blkback.max_buffer_pages` using > the `/sys/module/xen_blkback/parameters/max_buffer_pages` file. In this > test, I set the value to `1024` and `0`. The `1024` is the default > value. Setting the value as `0` is same to a situation doing the > squeezing always (worst-case). > > For the I/O performance measurement, I run a simple `dd` command 5 times > as below and collect the 'MB/s' results. > > $ for i in {1..5}; do dd if=/dev/zero of=file \ > bs=4k count=$((256*512)); sync; done I think it would be better if you could skip the filesystem overhead by writing directly to a block device, ie: Attach a null_blk based block device to the guest (on dom0): # xl block-attach guest phy:/dev/null_blk0 xvdb w Run a workload against the device (inside the guest): # fio --filename=/dev/xvdb --direct=1 --rw=randrw --bs=4k --ioengine=libaio \ --iodepth=64 --runtime=120 --numjobs=4 --time_based --group_reporting \ --name=throughput-test-job --eta-newline=1 You should run this on a multi-vcpu guest so that multiple queues are used, and adjust the numjobs to (at least) match the number of queues. > > If the underlying block device is slow enough, the squeezing overhead > could be hidden. For the reason, I do this test for both a slow block > device and a fast block device. I use a popular cloud block storage > service, ebs[1] as a slow device and the ramdisk block device[2] for the > fast device. > > The results are as below. 'max_pgs' represents the value of the > `blkback.max_buffer_pages` parameter. > > On the slow block device > ------------------------ > > max_pgs Min Max Median Avg Stddev > 0 38.7 45.8 38.7 40.12 3.1752165 > 1024 38.7 45.8 38.7 40.12 3.1752165 > No difference proven at 95.0% confidence > > On the fast block device > ------------------------ > > max_pgs Min Max Median Avg Stddev > 0 417 423 420 419.4 2.5099801 > 1024 414 425 416 417.8 4.4384682 > No difference proven at 95.0% confidence This is intriguing, as it seems to prove that the usage of a cache of free pages is irrelevant performance wise. The pool of free pages was introduced long ago, and it's possible that recent improvements to the balloon driver had made such pool useless, at which point it could be removed instead of worked around. Do you think you could perform some more tests (as pointed out above against the block device to skip the fs overhead) and report back the results? > > In short, even worst case squeezing on ramdisk based fast block device > makes no visible performance degradation. Please note that this is just > a very simple and minimal test. On systems using super-fast block > devices and a special I/O workload, the results might be different. If > you have any doubt, test on your machine for your workload to find the > optimal squeezing duration for you. > > [1] https://aws.amazon.com/ebs/ > [2] https://www.kernel.org/doc/html/latest/admin-guide/blockdev/ramdisk.html > > Reviewed-by: Juergen Gross <jgross@suse.com> > Signed-off-by: SeongJae Park <sjpark@amazon.de> > --- > drivers/block/xen-blkback/blkback.c | 22 ++++++++++++++++++++-- > drivers/block/xen-blkback/common.h | 1 + > drivers/block/xen-blkback/xenbus.c | 3 ++- > 3 files changed, 23 insertions(+), 3 deletions(-) > > diff --git a/drivers/block/xen-blkback/blkback.c b/drivers/block/xen-blkback/blkback.c > index fd1e19f1a49f..98823d150905 100644 > --- a/drivers/block/xen-blkback/blkback.c > +++ b/drivers/block/xen-blkback/blkback.c > @@ -142,6 +142,21 @@ static inline bool persistent_gnt_timeout(struct persistent_gnt *persistent_gnt) > HZ * xen_blkif_pgrant_timeout); > } > > +/* Once a memory pressure is detected, squeeze free page pools for a while. */ > +static unsigned int buffer_squeeze_duration_ms = 10; > +module_param_named(buffer_squeeze_duration_ms, > + buffer_squeeze_duration_ms, int, 0644); > +MODULE_PARM_DESC(buffer_squeeze_duration_ms, > +"Duration in ms to squeeze pages buffer when a memory pressure is detected"); You should add a description about this parameter to Documentation/ABI/testing/sysfs-driver-xen-blkback Thanks, Roger.
On Thu, 12 Dec 2019 12:42:47 +0100 "Roger Pau Monné" <roger.pau@citrix.com> wrote: > > Please make sure you Cc me in blkback related patches. Sorry for forgotting you! I will never forget again. > > On Wed, Dec 11, 2019 at 06:10:15PM +0000, SeongJae Park wrote: > > Each `blkif` has a free pages pool for the grant mapping. The size of > > the pool starts from zero and be increased on demand while processing > ^ is > > the I/O requests. If current I/O requests handling is finished or 100 > > milliseconds has passed since last I/O requests handling, it checks and > > shrinks the pool to not exceed the size limit, `max_buffer_pages`. > > > > Therefore, host administrators can cause memory pressure in blkback by > > attaching a large number of block devices and inducing I/O. Such > > problematic situations can be avoided by limiting the maximum number of > > devices that can be attached, but finding the optimal limit is not so > > easy. Improper set of the limit can results in the memory pressure or a > ^ s/the// > > resource underutilization. This commit avoids such problematic > > situations by squeezing the pools (returns every free page in the pool > > to the system) for a while (users can set this duration via a module > > parameter) if a memory pressure is detected. > ^ s/a// > > > > Discussions > > =========== > > > > The `blkback`'s original shrinking mechanism returns only pages in the > > pool, which are not currently be used by `blkback`, to the system. In > > I think you can remove both comas in the above sentence. > > > other words, the pages that are not mapped with granted pages. Because > > this commit is changing only the shrink limit but still uses the same > > freeing mechanism it does not touch pages which are currently mapping > > grants. > > > > Once a memory pressure is detected, this commit keeps the squeezing > ^ s/a// Thank you for corrections, will apply! > > limit for a user-specified time duration. The duration should be > > neither too long nor too short. If it is too long, the squeezing > > incurring overhead can reduce the I/O performance. If it is too short, > > `blkback` will not free enough pages to reduce the memory pressure. > > This commit sets the value as `10 milliseconds` by default because it is > > a short time in terms of I/O while it is a long time in terms of memory > > operations. Also, as the original shrinking mechanism works for at > > least every 100 milliseconds, this could be a somewhat reasonable > > choice. I also tested other durations (refer to the below section for > > more details) and confirmed that 10 milliseconds is the one that works > > best with the test. That said, the proper duration depends on actual > > configurations and workloads. That's why this commit allows users to > > set the duration as a module parameter. > > > > Memory Pressure Test > > ==================== > > > > To show how this commit fixes the memory pressure situation well, I > > configured a test environment on a xen-running virtualization system. > > On the `blkfront` running guest instances, I attach a large number of > > network-backed volume devices and induce I/O to those. Meanwhile, I > > measure the number of pages that swapped in (pswpin) and out (pswpout) > > on the `blkback` running guest. The test ran twice, once for the > > `blkback` before this commit and once for that after this commit. As > > shown below, this commit has dramatically reduced the memory pressure: > > > > pswpin pswpout > > before 76,672 185,799 > > after 212 3,325 > > > > Optimal Aggressive Shrinking Duration > > ------------------------------------- > > > > To find a best squeezing duration, I repeated the test with three > > different durations (1ms, 10ms, and 100ms). The results are as below: > > > > duration pswpin pswpout > > 1 852 6,424 > > 10 212 3,325 > > 100 203 3,340 > > > > As expected, the memory pressure has decreased as the duration is > > increased, but the reduction stopped from the `10ms`. Based on this > > results, I chose the default duration as 10ms. > > > > Performance Overhead Test > > ========================= > > > > This commit could incur I/O performance degradation under severe memory > > pressure because the squeezing will require more page allocations per > > I/O. To show the overhead, I artificially made a worst-case squeezing > > situation and measured the I/O performance of a `blkfront` running > > guest. > > > > For the artificial squeezing, I set the `blkback.max_buffer_pages` using > > the `/sys/module/xen_blkback/parameters/max_buffer_pages` file. In this > > test, I set the value to `1024` and `0`. The `1024` is the default > > value. Setting the value as `0` is same to a situation doing the > > squeezing always (worst-case). > > > > For the I/O performance measurement, I run a simple `dd` command 5 times > > as below and collect the 'MB/s' results. > > > > $ for i in {1..5}; do dd if=/dev/zero of=file \ > > bs=4k count=$((256*512)); sync; done > > I think it would be better if you could skip the filesystem overhead > by writing directly to a block device, ie: > > Attach a null_blk based block device to the guest (on dom0): > # xl block-attach guest phy:/dev/null_blk0 xvdb w > > Run a workload against the device (inside the guest): > # fio --filename=/dev/xvdb --direct=1 --rw=randrw --bs=4k --ioengine=libaio \ > --iodepth=64 --runtime=120 --numjobs=4 --time_based --group_reporting \ > --name=throughput-test-job --eta-newline=1 > > You should run this on a multi-vcpu guest so that multiple queues are > used, and adjust the numjobs to (at least) match the number of queues. I forgot to update the `dd` command. I used the command for the slow block device test, but directly induced the I/O towards the block device for the fast block device test as below: # xl block-attach guest phy:/dev/ram0 xvdb w $ for i in {1..5}; do dd if=/dev/zero of=file \ bs=4k count=$((256*512)); sync; done Nevertheless, I agree that you suggested test will provide much more accurate. As stated before, my test is only designed for a minimal proof-of-concept. > > > > > > If the underlying block device is slow enough, the squeezing overhead > > could be hidden. For the reason, I do this test for both a slow block > > device and a fast block device. I use a popular cloud block storage > > service, ebs[1] as a slow device and the ramdisk block device[2] for the > > fast device. > > > > The results are as below. 'max_pgs' represents the value of the > > `blkback.max_buffer_pages` parameter. > > > > On the slow block device > > ------------------------ > > > > max_pgs Min Max Median Avg Stddev > > 0 38.7 45.8 38.7 40.12 3.1752165 > > 1024 38.7 45.8 38.7 40.12 3.1752165 > > No difference proven at 95.0% confidence > > > > On the fast block device > > ------------------------ > > > > max_pgs Min Max Median Avg Stddev > > 0 417 423 420 419.4 2.5099801 > > 1024 414 425 416 417.8 4.4384682 > > No difference proven at 95.0% confidence > > This is intriguing, as it seems to prove that the usage of a cache of > free pages is irrelevant performance wise. > > The pool of free pages was introduced long ago, and it's possible that > recent improvements to the balloon driver had made such pool useless, > at which point it could be removed instead of worked around. I guess the grant page allocation overhead in this test scenario is really small. In an absence of memory pressure, fragmentation, and NUMA imbalance, the latency of the page allocation ('get_page()') is very short, as it will success in the fast path. Few years ago, I once measured the page allocation latency on my machine. Roughly speaking, it was about 1us in best case, 100us in worst case, and 5us in average. Please keep in mind that the measurement was not designed and performed in serious way. Thus the results could have profile overhead in it, though. While keeping that in mind, let's simply believe the number and ignore the latency of the block layer, blkback itself (including the grant mapping), and anything else including context switch, cache miss, but the allocation. In other words, suppose that the grant page allocation is only one source of the overhead. It will be able to achieve 1 million IOPS (4KB * 1MIOPS = 4 GB/s) in the best case, 200 thousand IOPS (800 MB/s) in average, and 10 thousand IOPS (40 MB/s) in worst case. Based on this coarse calculation, I think the test results is reasonable. This also means that the effect of the blkback's free pages pool might be visible under page allocation fast path failure situation. Nevertheless, it would be also hard to measure that in micro level unless the measurement is well designed and controlled. > > Do you think you could perform some more tests (as pointed out above > against the block device to skip the fs overhead) and report back the > results? To be honest, I'm not sure whether additional tests are really necessary, because I think the `dd` test and the results explanation already makes some sense and provide the minimal proof of the concept. Also, this change is a fallback for the memory pressure situation, which is an error path in some point of view. Such errorneous situation might not happen frequently and if the situation is not solved in short time, something much worse (e.g., OOM kill of the user space xen control processes) than temporal I/O performance degradation could happen. Thus, I'm not sure whether such detailed performance measurement is necessary for this rare error handling change. The comment of 'xen_blkbk_unmap()' also says, "This could accumulate ops up to the batch size to reduce the number of hypercalls, but since this is only used in error paths there's no real need.". That said, if you still want me to do the test, I will gladly do it. Note that since I'm now traveling US without the power cable of my now discharged laptop which is the only way to connect to my test environments in my office, I would not be able to do the additional test quickly. I assume I could do the test and give you the result within a couple of week. If you want me to post next version with minor changes such as commit message update meanwhile, please just let me know. Again, if you want it, I will gladly do it :) > > > > > In short, even worst case squeezing on ramdisk based fast block device > > makes no visible performance degradation. Please note that this is just > > a very simple and minimal test. On systems using super-fast block > > devices and a special I/O workload, the results might be different. If > > you have any doubt, test on your machine for your workload to find the > > optimal squeezing duration for you. > > > > [1] https://aws.amazon.com/ebs/ > > [2] https://www.kernel.org/doc/html/latest/admin-guide/blockdev/ramdisk.html > > > > Reviewed-by: Juergen Gross <jgross@suse.com> > > Signed-off-by: SeongJae Park <sjpark@amazon.de> > > --- > > drivers/block/xen-blkback/blkback.c | 22 ++++++++++++++++++++-- > > drivers/block/xen-blkback/common.h | 1 + > > drivers/block/xen-blkback/xenbus.c | 3 ++- > > 3 files changed, 23 insertions(+), 3 deletions(-) > > > > diff --git a/drivers/block/xen-blkback/blkback.c b/drivers/block/xen-blkback/blkback.c > > index fd1e19f1a49f..98823d150905 100644 > > --- a/drivers/block/xen-blkback/blkback.c > > +++ b/drivers/block/xen-blkback/blkback.c > > @@ -142,6 +142,21 @@ static inline bool persistent_gnt_timeout(struct persistent_gnt *persistent_gnt) > > HZ * xen_blkif_pgrant_timeout); > > } > > > > +/* Once a memory pressure is detected, squeeze free page pools for a while. */ > > +static unsigned int buffer_squeeze_duration_ms = 10; > > +module_param_named(buffer_squeeze_duration_ms, > > + buffer_squeeze_duration_ms, int, 0644); > > +MODULE_PARM_DESC(buffer_squeeze_duration_ms, > > +"Duration in ms to squeeze pages buffer when a memory pressure is detected"); > > You should add a description about this parameter to > Documentation/ABI/testing/sysfs-driver-xen-blkback Good point, I will. Thanks, SeongJae Park > > Thanks, Roger. >
On Thu, Dec 12, 2019 at 02:39:05PM +0100, SeongJae Park wrote: > On Thu, 12 Dec 2019 12:42:47 +0100 "Roger Pau Monné" <roger.pau@citrix.com> wrote: > > > On the slow block device > > > ------------------------ > > > > > > max_pgs Min Max Median Avg Stddev > > > 0 38.7 45.8 38.7 40.12 3.1752165 > > > 1024 38.7 45.8 38.7 40.12 3.1752165 > > > No difference proven at 95.0% confidence > > > > > > On the fast block device > > > ------------------------ > > > > > > max_pgs Min Max Median Avg Stddev > > > 0 417 423 420 419.4 2.5099801 > > > 1024 414 425 416 417.8 4.4384682 > > > No difference proven at 95.0% confidence > > > > This is intriguing, as it seems to prove that the usage of a cache of > > free pages is irrelevant performance wise. > > > > The pool of free pages was introduced long ago, and it's possible that > > recent improvements to the balloon driver had made such pool useless, > > at which point it could be removed instead of worked around. > > I guess the grant page allocation overhead in this test scenario is really > small. In an absence of memory pressure, fragmentation, and NUMA imbalance, > the latency of the page allocation ('get_page()') is very short, as it will > success in the fast path. The allocation of the pool of free pages involves more than get_page, it uses gnttab_alloc_pages which in the worse case will allocate a page and balloon it out issuing one hypercall. > Few years ago, I once measured the page allocation latency on my machine. > Roughly speaking, it was about 1us in best case, 100us in worst case, and 5us > in average. Please keep in mind that the measurement was not designed and > performed in serious way. Thus the results could have profile overhead in it, > though. While keeping that in mind, let's simply believe the number and ignore > the latency of the block layer, blkback itself (including the grant > mapping), and anything else including context switch, cache miss, but the > allocation. In other words, suppose that the grant page allocation is only one > source of the overhead. It will be able to achieve 1 million IOPS (4KB * > 1MIOPS = 4 GB/s) in the best case, 200 thousand IOPS (800 MB/s) in average, and > 10 thousand IOPS (40 MB/s) in worst case. Based on this coarse calculation, I > think the test results is reasonable. > > This also means that the effect of the blkback's free pages pool might be > visible under page allocation fast path failure situation. Nevertheless, it > would be also hard to measure that in micro level unless the measurement is > well designed and controlled. > > > > > Do you think you could perform some more tests (as pointed out above > > against the block device to skip the fs overhead) and report back the > > results? > > To be honest, I'm not sure whether additional tests are really necessary, > because I think the `dd` test and the results explanation already makes some > sense and provide the minimal proof of the concept. Also, this change is a > fallback for the memory pressure situation, which is an error path in some > point of view. Such errorneous situation might not happen frequently and if > the situation is not solved in short time, something much worse (e.g., OOM kill > of the user space xen control processes) than temporal I/O performance > degradation could happen. Thus, I'm not sure whether such detailed performance > measurement is necessary for this rare error handling change. Right, my main concern is that we seem to be adding duck tape so things don't fall apart, but if such cache is really not beneficial from a performance PoV I would rather see it go away than adding more stuff to it in order to workaround corner cases like memory starvation. Anyway, I guess we can take such change, but long term we need to look into fixing grants to not use ballooned pages, and figure out if the blkback free page cache is really useful or not. Thanks, Roger.
On Wed, Dec 11, 2019 at 06:10:15PM +0000, SeongJae Park wrote: > diff --git a/drivers/block/xen-blkback/blkback.c b/drivers/block/xen-blkback/blkback.c > index fd1e19f1a49f..98823d150905 100644 > --- a/drivers/block/xen-blkback/blkback.c > +++ b/drivers/block/xen-blkback/blkback.c > @@ -142,6 +142,21 @@ static inline bool persistent_gnt_timeout(struct persistent_gnt *persistent_gnt) > HZ * xen_blkif_pgrant_timeout); > } > > +/* Once a memory pressure is detected, squeeze free page pools for a while. */ > +static unsigned int buffer_squeeze_duration_ms = 10; > +module_param_named(buffer_squeeze_duration_ms, > + buffer_squeeze_duration_ms, int, 0644); > +MODULE_PARM_DESC(buffer_squeeze_duration_ms, > +"Duration in ms to squeeze pages buffer when a memory pressure is detected"); > + > +static unsigned long buffer_squeeze_end; > + > +void xen_blkbk_reclaim_memory(struct xenbus_device *dev) > +{ > + buffer_squeeze_end = jiffies + > + msecs_to_jiffies(buffer_squeeze_duration_ms); I'm not sure this is fully correct. This function will be called for each blkback instance, but the timeout is stored in a global variable that's shared between all blkback instances. Shouldn't this timeout be stored in xen_blkif so each instance has it's own local variable? Or else in the case you have 1k blkback instances the timeout is certainly going to be longer than expected, because each call to xen_blkbk_reclaim_memory will move it forward. Thanks, Roger.
On Thu, 12 Dec 2019 16:27:57 +0100 "Roger Pau Monné" <roger.pau@citrix.com> wrote: > > diff --git a/drivers/block/xen-blkback/blkback.c b/drivers/block/xen-blkback/blkback.c > > index fd1e19f1a49f..98823d150905 100644 > > --- a/drivers/block/xen-blkback/blkback.c > > +++ b/drivers/block/xen-blkback/blkback.c > > @@ -142,6 +142,21 @@ static inline bool persistent_gnt_timeout(struct persistent_gnt *persistent_gnt) > > HZ * xen_blkif_pgrant_timeout); > > } > > > > +/* Once a memory pressure is detected, squeeze free page pools for a while. */ > > +static unsigned int buffer_squeeze_duration_ms = 10; > > +module_param_named(buffer_squeeze_duration_ms, > > + buffer_squeeze_duration_ms, int, 0644); > > +MODULE_PARM_DESC(buffer_squeeze_duration_ms, > > +"Duration in ms to squeeze pages buffer when a memory pressure is detected"); > > + > > +static unsigned long buffer_squeeze_end; > > + > > +void xen_blkbk_reclaim_memory(struct xenbus_device *dev) > > +{ > > + buffer_squeeze_end = jiffies + > > + msecs_to_jiffies(buffer_squeeze_duration_ms); > > I'm not sure this is fully correct. This function will be called for > each blkback instance, but the timeout is stored in a global variable > that's shared between all blkback instances. Shouldn't this timeout be > stored in xen_blkif so each instance has it's own local variable? > > Or else in the case you have 1k blkback instances the timeout is > certainly going to be longer than expected, because each call to > xen_blkbk_reclaim_memory will move it forward. Agreed that. I think the extended timeout would not make a visible performance, though, because the time that 1k-loop take would be short enough to be ignored compared to the millisecond-scope duration. I took this way because I wanted to minimize such structural changes as far as I can, as this is just a point-fix rather than ultimate solution. That said, it is not fully correct and very confusing. My another colleague also pointed out it in internal review. Correct solution would be to adding a variable in the struct as you suggested or avoiding duplicated update of the variable by initializing the variable once the squeezing duration passes. I would prefer the later way, as it is more straightforward and still not introducing structural change. For example, it might be like below: diff --git a/drivers/block/xen-blkback/blkback.c b/drivers/block/xen-blkback/blkback.c index f41c698dd854..6856c8ef88de 100644 --- a/drivers/block/xen-blkback/blkback.c +++ b/drivers/block/xen-blkback/blkback.c @@ -152,8 +152,9 @@ static unsigned long buffer_squeeze_end; void xen_blkbk_reclaim_memory(struct xenbus_device *dev) { - buffer_squeeze_end = jiffies + - msecs_to_jiffies(buffer_squeeze_duration_ms); + if (!buffer_squeeze_end) + buffer_squeeze_end = jiffies + + msecs_to_jiffies(buffer_squeeze_duration_ms); } static inline int get_free_page(struct xen_blkif_ring *ring, struct page **page) @@ -669,10 +670,13 @@ int xen_blkif_schedule(void *arg) } /* Shrink the free pages pool if it is too large. */ - if (time_before(jiffies, buffer_squeeze_end)) + if (time_before(jiffies, buffer_squeeze_end)) { shrink_free_pagepool(ring, 0); - else + } else { + if (unlikely(buffer_squeeze_end)) + buffer_squeeze_end = 0; shrink_free_pagepool(ring, max_buffer_pages); + } if (log_stats && time_after(jiffies, ring->st_print)) print_stats(ring); May I ask you what way would you prefer? Thanks, SeongJae Park > > Thanks, Roger. >
On Thu, Dec 12, 2019 at 05:06:58PM +0100, SeongJae Park wrote: > On Thu, 12 Dec 2019 16:27:57 +0100 "Roger Pau Monné" <roger.pau@citrix.com> wrote: > > > > diff --git a/drivers/block/xen-blkback/blkback.c b/drivers/block/xen-blkback/blkback.c > > > index fd1e19f1a49f..98823d150905 100644 > > > --- a/drivers/block/xen-blkback/blkback.c > > > +++ b/drivers/block/xen-blkback/blkback.c > > > @@ -142,6 +142,21 @@ static inline bool persistent_gnt_timeout(struct persistent_gnt *persistent_gnt) > > > HZ * xen_blkif_pgrant_timeout); > > > } > > > > > > +/* Once a memory pressure is detected, squeeze free page pools for a while. */ > > > +static unsigned int buffer_squeeze_duration_ms = 10; > > > +module_param_named(buffer_squeeze_duration_ms, > > > + buffer_squeeze_duration_ms, int, 0644); > > > +MODULE_PARM_DESC(buffer_squeeze_duration_ms, > > > +"Duration in ms to squeeze pages buffer when a memory pressure is detected"); > > > + > > > +static unsigned long buffer_squeeze_end; > > > + > > > +void xen_blkbk_reclaim_memory(struct xenbus_device *dev) > > > +{ > > > + buffer_squeeze_end = jiffies + > > > + msecs_to_jiffies(buffer_squeeze_duration_ms); > > > > I'm not sure this is fully correct. This function will be called for > > each blkback instance, but the timeout is stored in a global variable > > that's shared between all blkback instances. Shouldn't this timeout be > > stored in xen_blkif so each instance has it's own local variable? > > > > Or else in the case you have 1k blkback instances the timeout is > > certainly going to be longer than expected, because each call to > > xen_blkbk_reclaim_memory will move it forward. > > Agreed that. I think the extended timeout would not make a visible > performance, though, because the time that 1k-loop take would be short enough > to be ignored compared to the millisecond-scope duration. > > I took this way because I wanted to minimize such structural changes as far as > I can, as this is just a point-fix rather than ultimate solution. That said, > it is not fully correct and very confusing. My another colleague also pointed > out it in internal review. Correct solution would be to adding a variable in > the struct as you suggested or avoiding duplicated update of the variable by > initializing the variable once the squeezing duration passes. I would prefer > the later way, as it is more straightforward and still not introducing > structural change. For example, it might be like below: > > diff --git a/drivers/block/xen-blkback/blkback.c b/drivers/block/xen-blkback/blkback.c > index f41c698dd854..6856c8ef88de 100644 > --- a/drivers/block/xen-blkback/blkback.c > +++ b/drivers/block/xen-blkback/blkback.c > @@ -152,8 +152,9 @@ static unsigned long buffer_squeeze_end; > > void xen_blkbk_reclaim_memory(struct xenbus_device *dev) > { > - buffer_squeeze_end = jiffies + > - msecs_to_jiffies(buffer_squeeze_duration_ms); > + if (!buffer_squeeze_end) > + buffer_squeeze_end = jiffies + > + msecs_to_jiffies(buffer_squeeze_duration_ms); > } > > static inline int get_free_page(struct xen_blkif_ring *ring, struct page **page) > @@ -669,10 +670,13 @@ int xen_blkif_schedule(void *arg) > } > > /* Shrink the free pages pool if it is too large. */ > - if (time_before(jiffies, buffer_squeeze_end)) > + if (time_before(jiffies, buffer_squeeze_end)) { > shrink_free_pagepool(ring, 0); > - else > + } else { > + if (unlikely(buffer_squeeze_end)) > + buffer_squeeze_end = 0; > shrink_free_pagepool(ring, max_buffer_pages); > + } > > if (log_stats && time_after(jiffies, ring->st_print)) > print_stats(ring); > > May I ask you what way would you prefer? I'm not particularly found of this approach, as I think it's racy. Ie: you would have to add some kind of lock to make sure the contents of buffer_squeeze_end stay unmodified during the read and set cycle, or else xen_blkif_schedule will race with xen_blkbk_reclaim_memory. This is likely not a big deal ATM since the code will work as expected in most cases AFAICT, but I would still prefer to have a per-instance buffer_squeeze_end added to xen_blkif, given that the callback is per-instance. I wouldn't call it a structural change, it's just adding a variable to a struct instead of having a shared one, but the code is almost the same as the current version. Thanks, Roger.
On 13.12.19 10:27, Roger Pau Monné wrote: > On Thu, Dec 12, 2019 at 05:06:58PM +0100, SeongJae Park wrote: >> On Thu, 12 Dec 2019 16:27:57 +0100 "Roger Pau Monné" <roger.pau@citrix.com> wrote: >> >>>> diff --git a/drivers/block/xen-blkback/blkback.c b/drivers/block/xen-blkback/blkback.c >>>> index fd1e19f1a49f..98823d150905 100644 >>>> --- a/drivers/block/xen-blkback/blkback.c >>>> +++ b/drivers/block/xen-blkback/blkback.c >>>> @@ -142,6 +142,21 @@ static inline bool persistent_gnt_timeout(struct persistent_gnt *persistent_gnt) >>>> HZ * xen_blkif_pgrant_timeout); >>>> } >>>> >>>> +/* Once a memory pressure is detected, squeeze free page pools for a while. */ >>>> +static unsigned int buffer_squeeze_duration_ms = 10; >>>> +module_param_named(buffer_squeeze_duration_ms, >>>> + buffer_squeeze_duration_ms, int, 0644); >>>> +MODULE_PARM_DESC(buffer_squeeze_duration_ms, >>>> +"Duration in ms to squeeze pages buffer when a memory pressure is detected"); >>>> + >>>> +static unsigned long buffer_squeeze_end; >>>> + >>>> +void xen_blkbk_reclaim_memory(struct xenbus_device *dev) >>>> +{ >>>> + buffer_squeeze_end = jiffies + >>>> + msecs_to_jiffies(buffer_squeeze_duration_ms); >>> >>> I'm not sure this is fully correct. This function will be called for >>> each blkback instance, but the timeout is stored in a global variable >>> that's shared between all blkback instances. Shouldn't this timeout be >>> stored in xen_blkif so each instance has it's own local variable? >>> >>> Or else in the case you have 1k blkback instances the timeout is >>> certainly going to be longer than expected, because each call to >>> xen_blkbk_reclaim_memory will move it forward. >> >> Agreed that. I think the extended timeout would not make a visible >> performance, though, because the time that 1k-loop take would be short enough >> to be ignored compared to the millisecond-scope duration. >> >> I took this way because I wanted to minimize such structural changes as far as >> I can, as this is just a point-fix rather than ultimate solution. That said, >> it is not fully correct and very confusing. My another colleague also pointed >> out it in internal review. Correct solution would be to adding a variable in >> the struct as you suggested or avoiding duplicated update of the variable by >> initializing the variable once the squeezing duration passes. I would prefer >> the later way, as it is more straightforward and still not introducing >> structural change. For example, it might be like below: >> >> diff --git a/drivers/block/xen-blkback/blkback.c b/drivers/block/xen-blkback/blkback.c >> index f41c698dd854..6856c8ef88de 100644 >> --- a/drivers/block/xen-blkback/blkback.c >> +++ b/drivers/block/xen-blkback/blkback.c >> @@ -152,8 +152,9 @@ static unsigned long buffer_squeeze_end; >> >> void xen_blkbk_reclaim_memory(struct xenbus_device *dev) >> { >> - buffer_squeeze_end = jiffies + >> - msecs_to_jiffies(buffer_squeeze_duration_ms); >> + if (!buffer_squeeze_end) >> + buffer_squeeze_end = jiffies + >> + msecs_to_jiffies(buffer_squeeze_duration_ms); >> } >> >> static inline int get_free_page(struct xen_blkif_ring *ring, struct page **page) >> @@ -669,10 +670,13 @@ int xen_blkif_schedule(void *arg) >> } >> >> /* Shrink the free pages pool if it is too large. */ >> - if (time_before(jiffies, buffer_squeeze_end)) >> + if (time_before(jiffies, buffer_squeeze_end)) { >> shrink_free_pagepool(ring, 0); >> - else >> + } else { >> + if (unlikely(buffer_squeeze_end)) >> + buffer_squeeze_end = 0; >> shrink_free_pagepool(ring, max_buffer_pages); >> + } >> >> if (log_stats && time_after(jiffies, ring->st_print)) >> print_stats(ring); >> >> May I ask you what way would you prefer? > > I'm not particularly found of this approach, as I think it's racy. Ie: > you would have to add some kind of lock to make sure the contents of > buffer_squeeze_end stay unmodified during the read and set cycle, or > else xen_blkif_schedule will race with xen_blkbk_reclaim_memory. > > This is likely not a big deal ATM since the code will work as > expected in most cases AFAICT, but I would still prefer to have a > per-instance buffer_squeeze_end added to xen_blkif, given that the > callback is per-instance. I wouldn't call it a structural change, it's > just adding a variable to a struct instead of having a shared one, but > the code is almost the same as the current version. FWIW, I agree. Juergen
On Fri, Dec 13, 2019 at 10:33 AM Jürgen Groß <jgross@suse.com> wrote: > > On 13.12.19 10:27, Roger Pau Monné wrote: > > On Thu, Dec 12, 2019 at 05:06:58PM +0100, SeongJae Park wrote: > >> On Thu, 12 Dec 2019 16:27:57 +0100 "Roger Pau Monné" <roger.pau@citrix.com> wrote: > >> > >>>> diff --git a/drivers/block/xen-blkback/blkback.c b/drivers/block/xen-blkback/blkback.c > >>>> index fd1e19f1a49f..98823d150905 100644 > >>>> --- a/drivers/block/xen-blkback/blkback.c > >>>> +++ b/drivers/block/xen-blkback/blkback.c > >>>> @@ -142,6 +142,21 @@ static inline bool persistent_gnt_timeout(struct persistent_gnt *persistent_gnt) > >>>> HZ * xen_blkif_pgrant_timeout); > >>>> } > >>>> > >>>> +/* Once a memory pressure is detected, squeeze free page pools for a while. */ > >>>> +static unsigned int buffer_squeeze_duration_ms = 10; > >>>> +module_param_named(buffer_squeeze_duration_ms, > >>>> + buffer_squeeze_duration_ms, int, 0644); > >>>> +MODULE_PARM_DESC(buffer_squeeze_duration_ms, > >>>> +"Duration in ms to squeeze pages buffer when a memory pressure is detected"); > >>>> + > >>>> +static unsigned long buffer_squeeze_end; > >>>> + > >>>> +void xen_blkbk_reclaim_memory(struct xenbus_device *dev) > >>>> +{ > >>>> + buffer_squeeze_end = jiffies + > >>>> + msecs_to_jiffies(buffer_squeeze_duration_ms); > >>> > >>> I'm not sure this is fully correct. This function will be called for > >>> each blkback instance, but the timeout is stored in a global variable > >>> that's shared between all blkback instances. Shouldn't this timeout be > >>> stored in xen_blkif so each instance has it's own local variable? > >>> > >>> Or else in the case you have 1k blkback instances the timeout is > >>> certainly going to be longer than expected, because each call to > >>> xen_blkbk_reclaim_memory will move it forward. > >> > >> Agreed that. I think the extended timeout would not make a visible > >> performance, though, because the time that 1k-loop take would be short enough > >> to be ignored compared to the millisecond-scope duration. > >> > >> I took this way because I wanted to minimize such structural changes as far as > >> I can, as this is just a point-fix rather than ultimate solution. That said, > >> it is not fully correct and very confusing. My another colleague also pointed > >> out it in internal review. Correct solution would be to adding a variable in > >> the struct as you suggested or avoiding duplicated update of the variable by > >> initializing the variable once the squeezing duration passes. I would prefer > >> the later way, as it is more straightforward and still not introducing > >> structural change. For example, it might be like below: > >> > >> diff --git a/drivers/block/xen-blkback/blkback.c b/drivers/block/xen-blkback/blkback.c > >> index f41c698dd854..6856c8ef88de 100644 > >> --- a/drivers/block/xen-blkback/blkback.c > >> +++ b/drivers/block/xen-blkback/blkback.c > >> @@ -152,8 +152,9 @@ static unsigned long buffer_squeeze_end; > >> > >> void xen_blkbk_reclaim_memory(struct xenbus_device *dev) > >> { > >> - buffer_squeeze_end = jiffies + > >> - msecs_to_jiffies(buffer_squeeze_duration_ms); > >> + if (!buffer_squeeze_end) > >> + buffer_squeeze_end = jiffies + > >> + msecs_to_jiffies(buffer_squeeze_duration_ms); > >> } > >> > >> static inline int get_free_page(struct xen_blkif_ring *ring, struct page **page) > >> @@ -669,10 +670,13 @@ int xen_blkif_schedule(void *arg) > >> } > >> > >> /* Shrink the free pages pool if it is too large. */ > >> - if (time_before(jiffies, buffer_squeeze_end)) > >> + if (time_before(jiffies, buffer_squeeze_end)) { > >> shrink_free_pagepool(ring, 0); > >> - else > >> + } else { > >> + if (unlikely(buffer_squeeze_end)) > >> + buffer_squeeze_end = 0; > >> shrink_free_pagepool(ring, max_buffer_pages); > >> + } > >> > >> if (log_stats && time_after(jiffies, ring->st_print)) > >> print_stats(ring); > >> > >> May I ask you what way would you prefer? > > > > I'm not particularly found of this approach, as I think it's racy. Ie: > > you would have to add some kind of lock to make sure the contents of > > buffer_squeeze_end stay unmodified during the read and set cycle, or > > else xen_blkif_schedule will race with xen_blkbk_reclaim_memory. > > > > This is likely not a big deal ATM since the code will work as > > expected in most cases AFAICT, but I would still prefer to have a > > per-instance buffer_squeeze_end added to xen_blkif, given that the > > callback is per-instance. I wouldn't call it a structural change, it's > > just adding a variable to a struct instead of having a shared one, but > > the code is almost the same as the current version. > > FWIW, I agree. Agreed, will send v8 soon! Thanks, SeongJae Park > > > Juergen
diff --git a/drivers/block/xen-blkback/blkback.c b/drivers/block/xen-blkback/blkback.c index fd1e19f1a49f..98823d150905 100644 --- a/drivers/block/xen-blkback/blkback.c +++ b/drivers/block/xen-blkback/blkback.c @@ -142,6 +142,21 @@ static inline bool persistent_gnt_timeout(struct persistent_gnt *persistent_gnt) HZ * xen_blkif_pgrant_timeout); } +/* Once a memory pressure is detected, squeeze free page pools for a while. */ +static unsigned int buffer_squeeze_duration_ms = 10; +module_param_named(buffer_squeeze_duration_ms, + buffer_squeeze_duration_ms, int, 0644); +MODULE_PARM_DESC(buffer_squeeze_duration_ms, +"Duration in ms to squeeze pages buffer when a memory pressure is detected"); + +static unsigned long buffer_squeeze_end; + +void xen_blkbk_reclaim_memory(struct xenbus_device *dev) +{ + buffer_squeeze_end = jiffies + + msecs_to_jiffies(buffer_squeeze_duration_ms); +} + static inline int get_free_page(struct xen_blkif_ring *ring, struct page **page) { unsigned long flags; @@ -656,8 +671,11 @@ int xen_blkif_schedule(void *arg) ring->next_lru = jiffies + msecs_to_jiffies(LRU_INTERVAL); } - /* Shrink if we have more than xen_blkif_max_buffer_pages */ - shrink_free_pagepool(ring, xen_blkif_max_buffer_pages); + /* Shrink the free pages pool if it is too large. */ + if (time_before(jiffies, buffer_squeeze_end)) + shrink_free_pagepool(ring, 0); + else + shrink_free_pagepool(ring, xen_blkif_max_buffer_pages); if (log_stats && time_after(jiffies, ring->st_print)) print_stats(ring); diff --git a/drivers/block/xen-blkback/common.h b/drivers/block/xen-blkback/common.h index 1d3002d773f7..1e0df86cb941 100644 --- a/drivers/block/xen-blkback/common.h +++ b/drivers/block/xen-blkback/common.h @@ -383,6 +383,7 @@ irqreturn_t xen_blkif_be_int(int irq, void *dev_id); int xen_blkif_schedule(void *arg); int xen_blkif_purge_persistent(void *arg); void xen_blkbk_free_caches(struct xen_blkif_ring *ring); +void xen_blkbk_reclaim_memory(struct xenbus_device *dev); int xen_blkbk_flush_diskcache(struct xenbus_transaction xbt, struct backend_info *be, int state); diff --git a/drivers/block/xen-blkback/xenbus.c b/drivers/block/xen-blkback/xenbus.c index b90dbcd99c03..0477f910b018 100644 --- a/drivers/block/xen-blkback/xenbus.c +++ b/drivers/block/xen-blkback/xenbus.c @@ -1115,7 +1115,8 @@ static struct xenbus_driver xen_blkbk_driver = { .ids = xen_blkbk_ids, .probe = xen_blkbk_probe, .remove = xen_blkbk_remove, - .otherend_changed = frontend_changed + .otherend_changed = frontend_changed, + .reclaim_memory = xen_blkbk_reclaim_memory, }; int xen_blkif_xenbus_init(void)
Each `blkif` has a free pages pool for the grant mapping. The size of the pool starts from zero and be increased on demand while processing the I/O requests. If current I/O requests handling is finished or 100 milliseconds has passed since last I/O requests handling, it checks and shrinks the pool to not exceed the size limit, `max_buffer_pages`. Therefore, host administrators can cause memory pressure in blkback by attaching a large number of block devices and inducing I/O. Such problematic situations can be avoided by limiting the maximum number of devices that can be attached, but finding the optimal limit is not so easy. Improper set of the limit can results in the memory pressure or a resource underutilization. This commit avoids such problematic situations by squeezing the pools (returns every free page in the pool to the system) for a while (users can set this duration via a module parameter) if a memory pressure is detected. Discussions =========== The `blkback`'s original shrinking mechanism returns only pages in the pool, which are not currently be used by `blkback`, to the system. In other words, the pages that are not mapped with granted pages. Because this commit is changing only the shrink limit but still uses the same freeing mechanism it does not touch pages which are currently mapping grants. Once a memory pressure is detected, this commit keeps the squeezing limit for a user-specified time duration. The duration should be neither too long nor too short. If it is too long, the squeezing incurring overhead can reduce the I/O performance. If it is too short, `blkback` will not free enough pages to reduce the memory pressure. This commit sets the value as `10 milliseconds` by default because it is a short time in terms of I/O while it is a long time in terms of memory operations. Also, as the original shrinking mechanism works for at least every 100 milliseconds, this could be a somewhat reasonable choice. I also tested other durations (refer to the below section for more details) and confirmed that 10 milliseconds is the one that works best with the test. That said, the proper duration depends on actual configurations and workloads. That's why this commit allows users to set the duration as a module parameter. Memory Pressure Test ==================== To show how this commit fixes the memory pressure situation well, I configured a test environment on a xen-running virtualization system. On the `blkfront` running guest instances, I attach a large number of network-backed volume devices and induce I/O to those. Meanwhile, I measure the number of pages that swapped in (pswpin) and out (pswpout) on the `blkback` running guest. The test ran twice, once for the `blkback` before this commit and once for that after this commit. As shown below, this commit has dramatically reduced the memory pressure: pswpin pswpout before 76,672 185,799 after 212 3,325 Optimal Aggressive Shrinking Duration ------------------------------------- To find a best squeezing duration, I repeated the test with three different durations (1ms, 10ms, and 100ms). The results are as below: duration pswpin pswpout 1 852 6,424 10 212 3,325 100 203 3,340 As expected, the memory pressure has decreased as the duration is increased, but the reduction stopped from the `10ms`. Based on this results, I chose the default duration as 10ms. Performance Overhead Test ========================= This commit could incur I/O performance degradation under severe memory pressure because the squeezing will require more page allocations per I/O. To show the overhead, I artificially made a worst-case squeezing situation and measured the I/O performance of a `blkfront` running guest. For the artificial squeezing, I set the `blkback.max_buffer_pages` using the `/sys/module/xen_blkback/parameters/max_buffer_pages` file. In this test, I set the value to `1024` and `0`. The `1024` is the default value. Setting the value as `0` is same to a situation doing the squeezing always (worst-case). For the I/O performance measurement, I run a simple `dd` command 5 times as below and collect the 'MB/s' results. $ for i in {1..5}; do dd if=/dev/zero of=file \ bs=4k count=$((256*512)); sync; done If the underlying block device is slow enough, the squeezing overhead could be hidden. For the reason, I do this test for both a slow block device and a fast block device. I use a popular cloud block storage service, ebs[1] as a slow device and the ramdisk block device[2] for the fast device. The results are as below. 'max_pgs' represents the value of the `blkback.max_buffer_pages` parameter. On the slow block device ------------------------ max_pgs Min Max Median Avg Stddev 0 38.7 45.8 38.7 40.12 3.1752165 1024 38.7 45.8 38.7 40.12 3.1752165 No difference proven at 95.0% confidence On the fast block device ------------------------ max_pgs Min Max Median Avg Stddev 0 417 423 420 419.4 2.5099801 1024 414 425 416 417.8 4.4384682 No difference proven at 95.0% confidence In short, even worst case squeezing on ramdisk based fast block device makes no visible performance degradation. Please note that this is just a very simple and minimal test. On systems using super-fast block devices and a special I/O workload, the results might be different. If you have any doubt, test on your machine for your workload to find the optimal squeezing duration for you. [1] https://aws.amazon.com/ebs/ [2] https://www.kernel.org/doc/html/latest/admin-guide/blockdev/ramdisk.html Reviewed-by: Juergen Gross <jgross@suse.com> Signed-off-by: SeongJae Park <sjpark@amazon.de> --- drivers/block/xen-blkback/blkback.c | 22 ++++++++++++++++++++-- drivers/block/xen-blkback/common.h | 1 + drivers/block/xen-blkback/xenbus.c | 3 ++- 3 files changed, 23 insertions(+), 3 deletions(-)