On Sun, Oct 22, 2017 at 11:27 PM, Xiaoxi Chen <superdebuger@xxxxxxxxx> wrote:
> To add another data point, switched to ceph-fuse 12.2.0, still seeing
> lots of lookup.
> lookup avg 1892
> mkdir avg 367
> create avg 222
> open avg 228
>
But in your test, mkdir avg was about 1.5 times of open avg. I think
your test created millions of directories, lookups were from cache
miss. You can try enlarging client_cache_size. But I don't think it
will help much when active set of directory are so large.
> 2017-10-21 2:09 GMT+08:00 Xiaoxi Chen <superdebuger@xxxxxxxxx>:
>> @Zheng, my kernel doesn't even has c3f4688a08f. But 200fd27 ("ceph:
>> use lookup request to revalidate dentry") is there.
>>
>> 2017-10-21 0:54 GMT+08:00 Xiaoxi Chen <superdebuger@xxxxxxxxx>:
>>> Thanks, will check.
>>>
>>> A general question, does cephfs kernel client drop dentries/inode
>>> cache aggressively? What I know is if MDS issue
>>> CEPH_SESSION_RECALL_STATE, client will drop, but is there other cases
>>> client will drop cache?
>>>
>>>
>>>
>>> 2017-10-20 16:39 GMT+08:00 Yan, Zheng <ukernel@xxxxxxxxx>:
>>>> On Fri, Oct 20, 2017 at 3:28 PM, Xiaoxi Chen <superdebuger@xxxxxxxxx> wrote:
>>>>> Centos 7.3, kernel version 3.10.0-514.26.2.el7.x86_64.
>>>>>
>>>>> I extract the logical of file creation in our workload into a
>>>>> reproducer , like below.
>>>>>
>>>>> Concurrently run the reproducer in 2+ node can see a lots of lookup OP.
>>>>> I thought the lookup is to open the directory tree so I tried to
>>>>> pre-make most of the dirs , use ls -i trying to read the dentries and
>>>>> cache it, then re-run the reproducer, seems nothing different..
>>>>>
>>>>> #include <sys/stat.h>
>>>>> #include <fcntl.h>
>>>>> int create_file(char * base, int count, int max, int depth)
>>>>> {
>>>>> int i;
>>>>> for(i=0; i<count; i++) {
>>>>> char dir[256];
>>>>> int mydir = rand() % max;
>>>>> sprintf(dir, "%s/%d", path, mydir);
>>>>> if (depth >=1) {
>>>>> mkdir(dir,0777);
>>>>> create_dir(dir, count, max, depth - 1);
>>>>> } else {
>>>>> int fd = open(dir, O_CREAT | O_EXCL| O_WRONLY , 0666);
>>>>> printf("opened path : %s = %d\n", path, fd);
>>>>> close(fd);
>>>>> }
>>>>> }
>>>>> }
>>>>> int main(int argc, char argv[])
>>>>> {
>>>>> char path[256];
>>>>> while(1) {
>>>>> create_file("/import/SQL01", 1, 4 ,10);
>>>>> }
>>>>> }
>>>>>
>>>>
>>>> still don't see this behavior on 4.13 kernel. I suspect there is
>>>> something wrong with dentry lease. please check if your kernel
>>>> include:
>>>>
>>>> commit c3f4688a08f (ceph: don't set req->r_locked_dir in ceph_d_revalidate)
>>>> commit 5eb9f6040f3 (ceph: do a LOOKUP in d_revalidate instead of GETATTR)
>>>>
>>>> The first commit can cause this issue, the second one fixes it.
>>>>
>>>> Regards
>>>> Yan, Zheng
>>>>
>>>>>
>>>>>
>>>>> 2017-10-20 10:55 GMT+08:00 Yan, Zheng <ukernel@xxxxxxxxx>:
>>>>>> On Fri, Oct 20, 2017 at 12:49 AM, Xiaoxi Chen <superdebuger@xxxxxxxxx> wrote:
>>>>>>> Hi,
>>>>>>>
>>>>>>> I am seeing a lot of lookup request when doing recursive mkdir.
>>>>>>> The workload behavior is like:
>>>>>>> mkdir DIR0
>>>>>>> mkdir DIR0/DIR1
>>>>>>> mkdir DIR0/DIR1/DIR2
>>>>>>> ....
>>>>>>> mkdir DIR0/DIR1/DIR2......./DIR7
>>>>>>> create DIR0/DIR1/DIR2......./DIR7/FILE1
>>>>>>>
>>>>>>> and concurrently run on 50+ clients, the dir name in different
>>>>>>> clients may or maynot be the same.
>>>>>>>
>>>>>>> from the admin socket I was seeing ~50K create requests, but
>>>>>>> got 400K lookup requests. The lookup eat up most of the mds capability
>>>>>>> so file create is slow.
>>>>>>>
>>>>>>> Where is the lookup comes from and can we have anyway to
>>>>>>> optimize it out ?
>>>>>>>
>>>>>>
>>>>>> I don't see this behavior when running following commands in 4.13
>>>>>> kernel client and luminous version ceph-fuse. which client do you use?
>>>>>>
>>>>>> mkdir d1
>>>>>> mkdir d1/d2
>>>>>> mkdir d1/d2/d3
>>>>>> mkdir d1/d2/d3/d4/
>>>>>> mkdir d1/d2/d3/d4/d5
>>>>>> touch d1/d2/d3/d4/d5/f
>>>>>>
>>>>>>> Xiaoxi
--
To unsubscribe from this list: send the line "unsubscribe ceph-devel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
