Tuning Linux for high performance environments

[joseph]

NIC driver tuning for better interrupt handling (Intel's)
Enabling TCP offload (TOE)
Increase TCP/UDP buffers
Increasing tcp queue len size
Enable jumbo frames if suitable for your work load
Increase shared memory/message queue buffers
Switching to the SCHED_RT (real time) or if available SCHED_EFD (deadline)
Switching to Deadline I/O Scheduler
Using async/directio mount options if your work load tolerates it.
Using a disk allocation unit (DAU) size that matches your i/o subsystem writes.
Enabling read ahead cache
Setting processor affinity on single thread apps to avoid ctx switching overhead.

[Iceman]

Is there any reasons for these adjustings? Personally it depends on the service contexts.
E.g. If your system bottleneck is mostly about network, a TOE net cards would be much better for performance boosting.