[LU-11618] implement ladvise rpc_size for optimized performance

• Status: Open
• Project: Lustre
• Type: New Feature
• Priority: Minor
• Reporter: Li Xi
• Assignee: Li Xi
• Resolution: Unresolved

Description

Originally, the maximum RPC size for bulk I/O can be controlled by
per-OSC parameter max_pages_per_rpc. Whenever it is possible, the
OSC will do bulk I/O as large as max_pages_per_rpc for better
performance. Thus, changing the value of max_pages_per_rpc usually
affects the I/O performance a lot. However, due to the I/O pattern
difference, not all applications can get the best performance with
the same value of max_pages_per_rpc.

We want to add a new type of ladvise to enabling applications to
set different RPC sizes for different files. max_pages_per_rpc is
still the upper limit of the RPC size in all cases. And new
parameter default_pages_per_rpc has been added and its value is
the default RPC size. If a ladvise of rpc_size is given to a file,
the RPC size of the file will be changed according to the ladvise.
But the maximum RPC size will still limited by max_pages_per_rpc.

The RPC size of a file configured by ladivse is neither a global
attribute nor a persistent attribute. Each client may have
different RPC size for the same file. And the RPC size of the file
will change back to default_pages_per_rpc when the hint kept in
memory is lost due to memory shrinkage.

Issue Links

• is related to: LU-6179 Lock ahead - Request extent locks fro...

Comments

[Gerrit Updater]

Li Xi ([e-mail]) uploaded a new patch: https://review.whamcloud.com/33573
Subject: LU-11618 osc: implement ladvise rpc_size
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: ce387d26f369dec24496eaed4f9213b76bb3f47f

[Andreas Dilger]

In the past, the RPC size was also affected by the stripe size, so that applications could specify this on a per-file basis. Also, storing the stripe_size in the file is persistent, and does not require the application to be modified.

[Li Xi]

Thanks Andreas. Has this mechanism of stripe-based-RPC-size being changed? I can't find any related codes now.

We are thinking of using ladvise hint through MPI ROMIO like LU-6179. That would require some arguments when calling MPI run command, but doesn't require modification of the appliation. And we are tired of tuning the global RPC size parameter to get good performance.

 

[Li Xi]

Following is how to use it:

MGS
$ lctl conf_param lipe1-OST*.obdfilter.brw_size=16
$ lctl conf_param lipe1-OST*.osc.max_pages_per_rpc=16M

Client
$ mount -t lustre 10.0.1.148@tcp:10.0.1.149@tcp:/lipe1 /mnt/lustre_lipe1
$ cat /proc/fs/lustre/osc/lipe1-OST0000-osc-*/default_pages_per_rpc
4096
$ cat /proc/fs/lustre/osc/lipe1-OST0000-osc-*/max_pages_per_rpc
4096
$ lfs setstripe -c 1 -i 0 /mnt/lustre_lipe1/file
$ dd if=/dev/zero of=/mnt/lustre_lipe1/file bs=1048576 count=10
$ cat /proc/fs/lustre/osc/lipe1-OST0000-osc-*/rpc_stats
...
read write
pages per rpc rpcs % cum % | rpcs % cum %
1: 0 0 0 | 0 0 0
2: 0 0 0 | 0 0 0
4: 0 0 0 | 0 0 0
8: 0 0 0 | 0 0 0
16: 0 0 0 | 0 0 0
32: 0 0 0 | 0 0 0
64: 0 0 0 | 0 0 0
128: 0 0 0 | 0 0 0
256: 0 0 0 | 0 0 0
512: 0 0 0 | 0 0 0
1024: 0 0 0 | 0 0 0
2048: 0 0 0 | 0 0 0
4096: 0 0 0 | 1 100 100
...
$ echo > /proc/fs/lustre/osc/lipe1-OST0000-osc-*/rpc_stats
$ echo 256 > /proc/fs/lustre/osc/lipe1-OST0000-osc-*/default_pages_per_rpc
$ dd if=/dev/zero of=/mnt/lustre_lipe1/file bs=1048576 count=10
$ cat /proc/fs/lustre/osc/lipe1-OST0000-osc-*/rpc_stats
...
read write
pages per rpc rpcs % cum % | rpcs % cum %
1: 0 0 0 | 0 0 0
2: 0 0 0 | 0 0 0
4: 0 0 0 | 0 0 0
8: 0 0 0 | 0 0 0
16: 0 0 0 | 0 0 0
32: 0 0 0 | 0 0 0
64: 0 0 0 | 0 0 0
128: 0 0 0 | 0 0 0
256: 0 0 0 | 10 100 100
...
$ echo > /proc/fs/lustre/osc/lipe1-OST0000-osc-*/rpc_stats
$ lfs ladvise -a rpcsize -r 16M /mnt/lustre_lipe1/file
$ dd if=/dev/zero of=/mnt/lustre_lipe1/file bs=1048576 count=16
$ cat /proc/fs/lustre/osc/lipe1-OST0000-osc-*/rpc_stats
...
read write
pages per rpc rpcs % cum % | rpcs % cum %
1: 0 0 0 | 0 0 0
2: 0 0 0 | 0 0 0
4: 0 0 0 | 0 0 0
8: 0 0 0 | 0 0 0
16: 0 0 0 | 0 0 0
32: 0 0 0 | 0 0 0
64: 0 0 0 | 0 0 0
128: 0 0 0 | 0 0 0
256: 0 0 0 | 0 0 0
512: 0 0 0 | 0 0 0
1024: 0 0 0 | 0 0 0
2048: 0 0 0 | 0 0 0
4096: 0 0 0 | 1 100 100
...
$ echo > /proc/fs/lustre/osc/lipe1-OST0000-osc-*/rpc_stats
$ lfs ladvise -a rpcsize -r 4M /mnt/lustre_lipe1/file
$ dd if=/dev/zero of=/mnt/lustre_lipe1/file bs=1048576 count=16
$ cat /proc/fs/lustre/osc/lipe1-OST0000-osc-*/rpc_stats
...
read write
pages per rpc rpcs % cum % | rpcs % cum %
1: 0 0 0 | 0 0 0
2: 0 0 0 | 0 0 0
4: 0 0 0 | 0 0 0
8: 0 0 0 | 0 0 0
16: 0 0 0 | 0 0 0
32: 0 0 0 | 0 0 0
64: 0 0 0 | 0 0 0
128: 0 0 0 | 0 0 0
256: 0 0 0 | 0 0 0
512: 0 0 0 | 0 0 0
1024: 0 0 0 | 4 100 100
...