[Share the google doc here to avoid SPAM detection] Here is the new testing result with multiple threads fio testing: https://docs.google.com/document/d/1AmbIEa_2MhB9bqhC3rfga9tp7n9YX9PLn0jSUxscVW0/edit?usp=sharing On Fri, Jan 8, 2021 at 4:47 PM Dongdong Tao <dongdong.tao@xxxxxxxxxxxxx> wrote: > > Yeap, I will scale the testing for multiple threads with larger IO > depth, thanks for the suggestion! > > On Fri, Jan 8, 2021 at 4:40 PM Coly Li <colyli@xxxxxxx> wrote: > > > > On 1/8/21 4:30 PM, Dongdong Tao wrote: > > > Hi Coly, > > > > > > They are captured with the same time length, the meaning of the > > > timestamp and the time unit on the x-axis are different. > > > (Sorry, I should have clarified this right after the chart) > > > > > > For the latency chart: > > > The timestamp is the relative time since the beginning of the > > > benchmark, so the start timestamp is 0 and the unit is based on > > > millisecond > > > > > > For the dirty data and cache available percent chart: > > > The timestamp is the UNIX timestamp, the time unit is based on second, > > > I capture the stats every 5 seconds with the below script: > > > --- > > > #!/bin/sh > > > while true; do echo "`date +%s`, `cat > > > /sys/block/bcache0/bcache/dirty_data`, `cat > > > /sys/block/bcache0/bcache/cache/cache_available_percent`, `cat > > > /sys/block/bcache0/bcache/writeback_rate`" >> $1; sleep 5; done; > > > --- > > > > > > Unfortunately, I can't easily make them using the same timestamp, but > > > I guess I can try to convert the UNIX timestamp to the relative time > > > like the first one. > > > But If we ignore the value of the X-axis, we can still roughly > > > compare them by using the length of the X-axis since they have the > > > same time length, > > > and we can see that the Master's write start hitting the backing > > > device when the cache_available_percent dropped to around 30. > > > > Copied, thanks for the explanation. The chart for single thread with io > > depth 1 is convinced IMHO :-) > > > > One more question, the benchmark is about a single I/O thread with io > > depth 1, which is not typical condition for real workload. Do you have > > plan to test the latency and IOPS for multiple threads with larger I/O > > depth ? > > > > > > Thanks. > > > > > > Coly Li > > > > > > > > > > On Fri, Jan 8, 2021 at 12:06 PM Coly Li <colyli@xxxxxxx> wrote: > > >> > > >> On 1/7/21 10:55 PM, Dongdong Tao wrote: > > >>> Hi Coly, > > >>> > > >>> > > >>> Thanks for the reminder, I understand that the rate is only a hint of > > >>> the throughput, it’s a value to calculate the sleep time between each > > >>> round of keys writeback, the higher the rate, the shorter the sleep > > >>> time, most of the time this means the more dirty keys it can writeback > > >>> in a certain amount of time before the hard disk running out of speed. > > >>> > > >>> > > >>> Here is the testing data that run on a 400GB NVME + 1TB NVME HDD > > >>> > > >> > > >> Hi Dongdong, > > >> > > >> Nice charts :-) > > >> > > >>> Steps: > > >>> > > >>> 1. > > >>> > > >>> make-bcache -B <HDD> -C <NVME> --writeback > > >>> > > >>> 2. > > >>> > > >>> sudo fio --name=random-writers --filename=/dev/bcache0 > > >>> --ioengine=libaio --iodepth=1 --rw=randrw --blocksize=64k,8k > > >>> --direct=1 --numjobs=1 --write_lat_log=mix --log_avg_msec=10 > > >>>> The fio benchmark commands ran for about 20 hours. > > >>> > > >> > > >> The time lengths of first 3 charts are 7.000e+7, rested are 1.60930e+9. > > >> I guess the time length of the I/O latency chart is 1/100 of the rested. > > >> > > >> Can you also post the latency charts for 1.60930e+9 seconds? Then I can > > >> compare the latency with dirty data and available cache charts. > > >> > > >> > > >> Thanks. > > >> > > >> > > >> Coly Li > > >> > > >> > > >> > > >> > > >> > > >>> > > >>> Let’s have a look at the write latency first: > > >>> > > >>> Master: > > >>> > > >>> > > >>> > > >>> Master+the patch: > > >>> > > >>> Combine them together: > > >>> > > >>> Again, the latency (y-axis) is based on nano-second, x-axis is the > > >>> timestamp based on milli-second, as we can see the master latency is > > >>> obviously much higher than the one with my patch when the master bcache > > >>> hit the cutoff writeback sync, the master isn’t going to get out of this > > >>> cutoff writeback sync situation, This graph showed it already stuck at > > >>> the cutoff writeback sync for about 4 hours before I finish the testing, > > >>> it may still needs to stuck for days before it can get out this > > >>> situation itself. > > >>> > > >>> > > >>> Note that there are 1 million points for each , red represents master, > > >>> green represents mater+my patch. Most of them are overlapped with each > > >>> other, so it may look like this graph has more red points then green > > >>> after it hitting the cutoff, but simply it’s because the latency has > > >>> scaled to a bigger range which represents the HDD latency. > > >>> > > >>> > > >>> > > >>> Let’s also have a look at the bcache’s cache available percent and dirty > > >>> data percent. > > >>> > > >>> Master: > > >>> > > >>> Master+this patch: > > >>> > > >>> As you can see, this patch can avoid it hitting the cutoff writeback sync. > > >>> > > >>> > > >>> As to say the improvement for this patch against the first one, let’s > > >>> take a look at the writeback rate changing during the run. > > >>> > > >>> patch V1: > > >>> > > >>> > > >>> > > >>> Patch V2: > > >>> > > >>> > > >>> The Y-axis is the value of rate, the V1 is very aggressive as it jumps > > >>> instantly from a minimum 8 to around 10 million. And the patch V2 can > > >>> control the rate under 5000 during the run, and after the first round of > > >>> writeback, it can stay even under 2500, so this proves we don’t need to > > >>> be as aggressive as V1 to get out of the high fragment situation which > > >>> eventually causes all writes hitting the backing device. This looks very > > >>> reasonable for me now. > > >>> > > >>> Note that the fio command that I used is consuming the bucket quite > > >>> aggressively, so it had to hit the third stage which has the highest > > >>> aggressiveness, but I believe this is not true in a real production env, > > >>> real production env won’t consume buckets that aggressively, so I expect > > >>> stage 3 may not very often be needed to hit. > > >>> > > >>> > > >>> As discussed, I'll run multiple block size testing on at least 1TB NVME > > >>> device later. > > >>> But it might take some time. > > >>> > > >>> > > >>> Regards, > > >>> Dongdong > > >>> > > >>> On Tue, Jan 5, 2021 at 12:33 PM Coly Li <colyli@xxxxxxx > > >>> <mailto:colyli@xxxxxxx>> wrote: > > >>> > > >>> On 1/5/21 11:44 AM, Dongdong Tao wrote: > > >>> > Hey Coly, > > >>> > > > >>> > This is the second version of the patch, please allow me to explain a > > >>> > bit for this patch: > > >>> > > > >>> > We accelerate the rate in 3 stages with different aggressiveness, the > > >>> > first stage starts when dirty buckets percent reach above > > >>> > BCH_WRITEBACK_FRAGMENT_THRESHOLD_LOW(50), the second is > > >>> > BCH_WRITEBACK_FRAGMENT_THRESHOLD_MID(57) and the third is > > >>> > BCH_WRITEBACK_FRAGMENT_THRESHOLD_HIGH(64). By default the first stage > > >>> > tries to writeback the amount of dirty data in one bucket (on average) > > >>> > in (1 / (dirty_buckets_percent - 50)) second, the second stage > > >>> tries to > > >>> > writeback the amount of dirty data in one bucket in (1 / > > >>> > (dirty_buckets_percent - 57)) * 200 millisecond. The third stage tries > > >>> > to writeback the amount of dirty data in one bucket in (1 / > > >>> > (dirty_buckets_percent - 64)) * 20 millisecond. > > >>> > > > >>> > As we can see, there are two writeback aggressiveness increasing > > >>> > strategies, one strategy is with the increasing of the stage, the > > >>> first > > >>> > stage is the easy-going phase whose initial rate is trying to > > >>> write back > > >>> > dirty data of one bucket in 1 second, the second stage is a bit more > > >>> > aggressive, the initial rate tries to writeback the dirty data of one > > >>> > bucket in 200 ms, the last stage is even more, whose initial rate > > >>> tries > > >>> > to writeback the dirty data of one bucket in 20 ms. This makes sense, > > >>> > one reason is that if the preceding stage couldn’t get the > > >>> fragmentation > > >>> > to a fine stage, then the next stage should increase the > > >>> aggressiveness > > >>> > properly, also it is because the later stage is closer to the > > >>> > bch_cutoff_writeback_sync. Another aggressiveness increasing > > >>> strategy is > > >>> > with the increasing of dirty bucket percent within each stage, the > > >>> first > > >>> > strategy controls the initial writeback rate of each stage, while this > > >>> > one increases the rate based on the initial rate, which is > > >>> initial_rate > > >>> > * (dirty bucket percent - BCH_WRITEBACK_FRAGMENT_THRESHOLD_X). > > >>> > > > >>> > The initial rate can be controlled by 3 parameters > > >>> > writeback_rate_fp_term_low, writeback_rate_fp_term_mid, > > >>> > writeback_rate_fp_term_high, they are default 1, 5, 50, users can > > >>> adjust > > >>> > them based on their needs. > > >>> > > > >>> > The reason that I choose 50, 57, 64 as the threshold value is because > > >>> > the GC must be triggered at least once during each stage due to the > > >>> > “sectors_to_gc” being set to 1/16 (6.25 %) of the total cache > > >>> size. So, > > >>> > the hope is that the first and second stage can get us back to good > > >>> > shape in most situations by smoothly writing back the dirty data > > >>> without > > >>> > giving too much stress to the backing devices, but it might still > > >>> enter > > >>> > the third stage if the bucket consumption is very aggressive. > > >>> > > > >>> > This patch use (dirty / dirty_buckets) * fp_term to calculate the > > >>> rate, > > >>> > this formula means that we want to writeback (dirty / > > >>> dirty_buckets) in > > >>> > 1/fp_term second, fp_term is calculated by above aggressiveness > > >>> > controller, “dirty” is the current dirty sectors, “dirty_buckets” > > >>> is the > > >>> > current dirty buckets, so (dirty / dirty_buckets) means the average > > >>> > dirty sectors in one bucket, the value is between 0 to 1024 for the > > >>> > default setting, so this formula basically gives a hint that to > > >>> reclaim > > >>> > one bucket in 1/fp_term second. By using this semantic, we can have a > > >>> > lower writeback rate when the amount of dirty data is decreasing and > > >>> > overcome the fact that dirty buckets number is always increasing > > >>> unless > > >>> > GC happens. > > >>> > > > >>> > *Compare to the first patch: > > >>> > *The first patch is trying to write back all the data in 40 seconds, > > >>> > this will result in a very high writeback rate when the amount of > > >>> dirty > > >>> > data is big, this is mostly true for the large cache devices. The > > >>> basic > > >>> > problem is that the semantic of this patch is not ideal, because we > > >>> > don’t really need to writeback all dirty data in order to solve this > > >>> > issue, and the instant large increase of the rate is something I > > >>> feel we > > >>> > should better avoid (I like things to be smoothly changed unless no > > >>> > choice: )). > > >>> > > > >>> > Before I get to this new patch(which I believe should be optimal > > >>> for me > > >>> > atm), there have been many tuning/testing iterations, eg. I’ve > > >>> tried to > > >>> > tune the algorithm to writeback ⅓ of the dirty data in a certain > > >>> amount > > >>> > of seconds, writeback 1/fragment of the dirty data in a certain amount > > >>> > of seconds, writeback all the dirty data only in those error_buckets > > >>> > (error buckets = dirty buckets - 50% of the total buckets) in a > > >>> certain > > >>> > amount of time. However, those all turn out not to be ideal, only the > > >>> > semantic of the patch makes much sense for me and allows me to control > > >>> > the rate in a more precise way. > > >>> > > > >>> > *Testing data: > > >>> > *I'll provide the visualized testing data in the next couple of days > > >>> > with 1TB NVME devices cache but with HDD as backing device since it's > > >>> > what we mostly used in production env. > > >>> > I have the data for 400GB NVME, let me prepare it and take it for > > >>> you to > > >>> > review. > > >>> [snipped] > > >>> > > >>> Hi Dongdong, > > >>> > > >>> Thanks for the update and continuous effort on this idea. > > >>> > > >>> Please keep in mind the writeback rate is just a advice rate for the > > >>> writeback throughput, in real workload changing the writeback rate > > >>> number does not change writeback throughput obviously. > > >>> > > >>> Currently I feel this is an interesting and promising idea for your > > >>> patch, but I am not able to say whether it may take effect in real > > >>> workload, so we do need convinced performance data on real workload and > > >>> configuration. > > >>> > > >>> Of course I may also help on the benchmark, but my to-do list is long > > >>> enough and it may take a very long delay time. > > >>> > > >>> Thanks. > > >>> > > >>> Coly Li > > >>> > > >> > >
