[lustre-discuss] lustre write wrong data under postgresql benchmark tool test(concurrent access same data file with primary node write and standy node read)

[Andreas Dilger]

Is PG using O_DIRECT or buffered read/write?  Is it caching the pages in userspace?

Lustre will definitely keep pages consistent between clients, but if the application is caching the pages in userspace, and does not have any protocol between the nodes to invalidate cached pages when they are modified on disk, then the data will become inconsistent when one node modifies it.

That is the same reason it isn't possible to mount a single ext4 filesystem r/w on one node and r/o on another node with shared storage, because the filesystem doesn't expect data to be changing underneath it, and will cache pages in RAM and not re-read them if they are modified on the other node.

Cheers, Andreas

On Sep 29, 2018, at 07:57, 王亮 <wanziforever at gmail.com<mailto:wanziforever at gmail.com>> wrote:

Hello, lustre development team

background: we have two postgresql instances running as a primary and standby and they share the same xlog file and data files (we change PG code to achieve this) which located in the mounted lustre file system, and we want to have a try with the lustre file system, we used gfs2 before, and we expect the lustre will show a much better performance, but ...

We meet a read/write concurrent access problem related with Lustre. Would you like to give us some suggestions? Any advices are appreciated, and thank you in advanced : )
note: we are very sure the standy instance will not write any data to disk. (to be sure of this, we also shutdown the standby end, and use pg_xlogdump tool to read the xlog file, the problem still happened, and pg_xlogdump to is just a query to with any write operation)


Scenario Description:
There’re 4 nodes(CentOS Linux release 7.4) connected with infiniband network(driven by MLNX_OFED_LINUX-4.4):
10.0.0.106 acts as MDS with a local PCI-E 800GB SSD that used as MDT.
10.0.0.101 acts as OSS with a same local PCI-E 800GB SSD that used as OST.
10.0.0.104 and 10.0.0.105 act as Lustre client and mount the Lustre file system at the directory of “/lustre”.
The Lustre related packages are compiled from official lustre-2.10.5-1.src.rpm.

The simplest verification(i.e. dd command) passed without errors.

Error:
Then start our customized PostgreSQL service at 104 and 105. 104 runs as the primary PostgreSQL server, and 105 runs as the secondary PostgreSQL server. All the two PostgreSQL nodes read/write the shared directory of “/lustre” provided by Lustre. The primary PostgreSQL server will open files with *RW* mode and write something into the files; at the *meantime* the second PostgreSQL server will open the same files with *R* mode and read the data written by the primary PostgreSQL server, and it gets the *wrong* data (the flushed data by primary in disk is error i.e. write wrong data into disk). This will happen when we run a benchmark tool of PostgreSQL.


PS 1.
We tried different options to mount the Lustre:
mount -t lustre -o flock 10.0.0.106 at o2ib0:/birdfs<mailto:10.0.0.106 at o2ib0:/birdfs> /lustre
mount -t lustre -o flock -o ro 10.0.0.106 at o2ib0:/birdfs<mailto:10.0.0.106 at o2ib0:/birdfs> /lustre
but the error always exists.

PS 2.
Attach the initial information, maybe helpful.
[root at 106 ~]# mkfs.lustre --fsname=birdfs --mgs --mdt --index=0 --reformat /dev/nvme0n1

   Permanent disk data:
Target:     birdfs:MDT0000
Index:      0
Lustre FS:  birdfs
Mount type: ldiskfs
Flags:      0x65
              (MDT MGS first_time update )
Persistent mount opts: user_xattr,errors=remount-ro
Parameters:

device size = 763097MB
formatting backing filesystem ldiskfs on /dev/nvme0n1
       target name   birdfs:MDT0000
       4k blocks     195353046
       options        -J size=4096 -I 1024 -i 2560 -q -O dirdata,uninit_bg,^extents,dir_nlink,quota,huge_file,flex_bg -E lazy_journal_init -F
mkfs_cmd = mke2fs -j -b 4096 -L birdfs:MDT0000  -J size=4096 -I 1024 -i 2560 -q -O dirdata,uninit_bg,^extents,dir_nlink,quota,huge_file,flex_bg -E lazy_journal_init -F /dev/nvme0n1 195353046

Writing CONFIGS/mountdata

[root at 101 ~]# mkfs.lustre --fsname=birdfs --ost --reformat --index=0 --mgsnode=10.0.0.106 at o2ib0<mailto:--mgsnode=10.0.0.106 at o2ib0> /dev/nvme0n1

   Permanent disk data:
Target:     birdfs:OST0000
Index:      0
Lustre FS:  birdfs
Mount type: ldiskfs
Flags:      0x62
              (OST first_time update )
Persistent mount opts: ,errors=remount-ro
Parameters: mgsnode=10.0.0.106 at o2ib<mailto:mgsnode=10.0.0.106 at o2ib>

device size = 763097MB
formatting backing filesystem ldiskfs on /dev/nvme0n1
       target name   birdfs:OST0000
       4k blocks     195353046
       options        -J size=400 -I 512 -i 69905 -q -O extents,uninit_bg,dir_nlink,quota,huge_file,flex_bg -G 256 -E resize="4290772992",lazy_journal_init -F
mkfs_cmd = mke2fs -j -b 4096 -L birdfs:OST0000  -J size=400 -I 512 -i 69905 -q -O extents,uninit_bg,dir_nlink,quota,huge_file,flex_bg -G 256 -E resize="4290772992",lazy_journal_init -F /dev/nvme0n1 195353046
Writing CONFIGS/mountdata


Looking forward to any replies.

Regards,
Bird


--
regards
denny
_______________________________________________
lustre-discuss mailing list
lustre-discuss at lists.lustre.org<mailto:lustre-discuss at lists.lustre.org>
http://lists.lustre.org/listinfo.cgi/lustre-discuss-lustre.org
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.lustre.org/pipermail/lustre-discuss-lustre.org/attachments/20180929/cbca16f0/attachment-0001.html>