[HPDD-discuss] poor I/O performance with client 2.4.3 on sles11sp1

Hi all, we have a Lustre 2.4.3 with - oss/mds: CentOS 6.5, whamcloud patched kernel 2.6.32-358.23.2.el6_lustre.x86_64 (rpms from http://downloads.whamcloud.com/public/lustre/lustre-2.4.3/el6/server/RPMS... ) - client: sles11sp1, kernel 2.6.32.36-0.5-default, patchless client from http://build.whamcloud.com/job/lustre-b2_4/74/arch=x86_64,build_type=clie... We are experiencing problems with basic I/O commands like `dd` and `cp`. When we write with `dd`, from a client running SLES11 SP1, *without direct I/O*, we got about ~40MB/s throughput. When we try with direct I/O (dd oflag=direct), we have the expected performance (220-240MB/s thriughput). If we run the exact same tests using a CentOS 6.5 with whamcloud patched kernel 2.6.32-358.23.2.el6_lustre.x86_64, we got the correct performance regardless of direct-I/O or non-direct I/O . We guess there is a problem with the kernel version we are using on our clients. Is there something known about this behaviour? Thanks, Francesco