[LU-2001] read operation is slow when mmap is enabled

• Type: Bug
• Resolution: Won't Fix
• Priority: Minor
• Affects Version/s: Lustre 1.8.8
• Labels: patch
• Environment: lustre-1.8.8 CentOS5.8
• Severity: 3
• Assignee: Jinshan Xiong (Inactive)
• Reporter: Shuichi Ihara (Inactive)

Description

There is an application and running on the lustre.

When we run this application trough the lustre client, it's really slow and application seems to be stall.
However, if the lustre exports lustre filesystem with NFS and mount it client itself, then application does IOs to lustre through the NFS, the performance seems to be reasonable.

1. mount -t lustre
   192.168.100.131@tcp:/lustre on /lustre type lustre (rw)
2. exportfs
   /lustre <world>
3. mount -t nfs
   localhost:/lustre on /lustre-nfs type nfs (rw,addr=127.0.0.1)

The application is doing the following operations.
1. 4 jobs (same application) are running on the single client and these jobs read single shared file as input file. file is open with open() and mmap().
2. computes with them and each job write output file separately. the block size 8k for read() and write().

Any ideas why application is fast if it goes through the NFS layer to lustre?