Comments on: tkprof (1)

By: goran

goran — Fri, 13 Jul 2007 14:53:14 +0000

increaseing the default values for send and receive queue sizes associated with networking devices, also adjusting TCP socket buffer settings, jumbo frames etc could help in reducing the number of network round-trips if network is the main problem in this case…
would be interesting to check the code and the logic…

By: Jonathan Lewis

Jonathan Lewis — Tue, 03 Jul 2007 17:18:35 +0000

Jeroen, thanks for the information.

By: Jeroen

Jeroen — Tue, 03 Jul 2007 10:17:58 +0000

MTU controls the Ethernet frame size whereas tcp_xmit_hiwat and tcp_rec_hiwat are typically an in-memory kernel buffer used as a store/forward mechanism between the applicaiton and kernel driver (sort of like a filesystem buffer). In fact you brought up a good point: the MTU needs to be adjusted to support larger frames 9200 bytes (think it’s called Jumbo Frames) if one really wants to cut down on the number of network round-trips for delivering a fixed-size payload.

By: Jonathan Lewis

Jonathan Lewis — Fri, 29 Jun 2007 07:14:24 +0000

Stew, thanks for the information; your description certainly fits the 'tiny lag' observations. Am I right in thinking that the 'tcp buffer size' is not the same thing as the MTU (maximum transmission unit) size which gets reported by some versions of the netstat command ? My memory of that figure is that it's often about 1450 bytes.

By: Stew Ashton

Stew Ashton — Thu, 28 Jun 2007 18:17:24 +0000

Jonathan,

In one of my projects, a colleague did a network analysis between Oracle and our application. He found that the TCP buffer size (default 16K in our OS) was too small for the 100 Mbit throughput of the network adapter. He upped our “standard” TCP buffer size to 64K. This might explain these tiny lags you speak of.

I would also mention that there is a slight additional latency of 1 to 3 milliseconds if the two machines are separated by routers and / or firewalls. This is probably only a problem when one “transaction” is generating many many SQL statements.

I spent a week testing different network-related parameters (TDU and SDU size mostly) and found no measurable difference.