Designing a High Performance Network File Server

Checksum Offload is supported on “iether”, “igelan” and “gelan” type Gigabit Ethernet interfaces as

well as “hpixgbe” type 10 Gigabit Ethernet adapters. CKO may be enabled or disabled for both the

transmit side or receive side of each Ethernet card. Configuring CKO can either be done interactively

using the lanadmin(1M) command with the “-X send_cko_on” or “-X recv_cko_on” options

or automatically during system boot by configuring the appropriate configuration file for your NIC

type in the /etc/rc.config.d directory.

For example, to enable send and receive CKO on an “iether” type interface add lines similar to the

following to the /etc/rc.config.d/hpietherconf file:

HP_IETHER_SEND_CKO[1]=ON

HP_IETHER_RECV_CKO[1]=ON

Similarly, to enable CKO on an “igelan” interface, add appropriate entries for the specific interface

to the /etc/rc.config.d/hpigelanconf file:

HP_IGELAN_SEND_CKO[1]=ON

HP_IGELAN_RECV_CKO[1]=ON

To enable CKO on a “gelan” interface, add appropriate entries for the specific interface to the

/etc/rc.config.d/hpgelanconf file:

HP_GELAN_SEND_CKO[1]=ON

HP_GELAN_RECV_CKO[1]=ON

Finally, to enable CKO on an “hpixgbe” interface, add appropriate entries for the specific interface to

the /etc/rc.config.d/hpixgbeconf file:

HP_IXGBE_TX_CHECKSUM_OFFLOAD[1]=ON

HP_IXGBE_RX_CHECKSUM_OFFLOAD[1]=ON

Increase TCP Send and Receive Windows

TCP uses what is called the “congestion window” to determine how many packets can be sent or

received on a connection at any given time. A well-tuned congestion window size enables higher

throughput on a connection, provided both the client and server systems are able to keep pace with

the rate of inbound and outbound packets.

On HP-UX 11i v2, the maximum congestion window sizes are configurable for both the send and

receive ends of the socket. These buffer sizes are controlled by the ndd(1M) parameters

tcp_xmit_hiwater_def (send) and tcp_recv_hiwater_def (receive). By default these

parameters are set to 32 Kilobytes. During this engagement these parameters were increased to 256

Kilobytes on the NFS server and all of the NFS clients. The increase in TCP window size improved

overall TCP throughput, and thereby increased NFS throughput.

The ndd(1M) command may be used to dynamically change these send and receive buffer sizes

interactively. However, since the TCP window size is negotiated during connection establishment,

only new TCP connections established after the values are changed will take advantage of the new

buffer size values. HP-UX systems may also be configured to increase the TCP buffer sizes during

system boot by adding the following entries to the /etc/rc.config.d/nddconf file:

TRANSPORT_NAME[0]=tcp

NDD_NAME[0]=tcp_recv_hiwater_def

NDD_VALUE[0]=262144