Scott Feldman announced that with the release of development kernel 2.5.33, the e1000 driver now supports TCP Segmentation Offloading (TSO), offering a significant boost in two-way transfer rates. (In the provided benchmark, send only throughput did not increase as the wire's physicial limitation had already been reached.)

TCP Segmentation Offload (or TCP Large Send) is when buffer's much larger than the supported maximum transmission unit (MTU) of a given medium are passed through the bus to the network interface card. The work of dividing the much larger packets into smaller packets is thus offloaded to the NIC. More specifically, the e1000 driver is passing 64k packets to the network card, which then divides these into proper MTU-sized 1500 byte packets.

Alexey Kuznetsov added TSO support into the stack, noting that as of yet, "the implementation in tcp is still at [the] level of a toy". Be that as it may, it's a good start, and before long other TSO capable devices will likely also be supported.

From: Scott Feldman
To: linux-kernel mailing list, linux-net
Subject: TCP Segmentation Offloading (TSO)
Date: 	Mon, 2 Sep 2002 10:45:08 -0700 

TCP Segmentation Offloading (TSO) is enabled[1] in 2.5.33, along with an
enabled e1000 driver.  Other capable devices can be enabled ala e1000; the
driver interface (NETIF_F_TSO) is very simple.

So, fire up you favorite networking performance tool and compare the
performance gains between 2.5.32 and 2.5.33 using e1000.  I ran a quick test
on a dual P4 workstation system using the commercial tool Chariot:

Tx/Rx TCP file send long (bi-directional Rx/Tx)
  w/o TSO: 1500Mbps, 82% CPU
  w/  TSO: 1633Mbps, 75% CPU
Tx TCP file send long (Tx only)
  w/o TSO: 940Mbps, 40% CPU
  w/  TSO: 940Mbps, 19% CPU
A good bump in throughput for the bi-directional test.  The Tx-only test was

already at wire speed, so the gains are pure CPU savings.
I'd like to see SPECWeb results w/ and w/o TSO, and any other relevant

testing.  UDP framentation is not offloaded, so keep testing to TCP.
-scott
[1] Kudos to Alexey Kuznetsov for enabling the stack with TSO support, to

Chris Leech for providing the e1000 bits and a prototype stack, and to David

Miller for consultation.



From: David S. Miller

Subject: Re: TCP Segmentation Offloading (TSO)

Date: 	Mon, 02 Sep 2002 16:13:05 -0700 (PDT)
I would like to praise Intel for working so closely with us on

this.  They gave us immediately, in one email, all the information we

needed to implement and test e1000 support for TSO under Linux.
With some other companies, doing this is like pulling teeth.



From: Alexey Kuznetsov

Subject: Re: TCP Segmentation Offloading (TSO)

Date: 	Mon, 2 Sep 2002 22:58:15 +0400 (MSD)
Hello!
> [1] Kudos to 
Hmm... wait awhile with celebrating, the implementation in tcp is still

at level of a toy. Well, and it happens to crash, the patch is enclosed.
Alexey
[patch]



From: Jordi Ros

Subject: RE: TCP Segmentation Offloading (TSO)

Date: 	Mon, 2 Sep 2002 21:58:32 -0700
One question regarding the throughput numbers,
what was the size of the packets built at the tcp layer (mss)?
i assume the mtu is ethernet 1500 Bytes, right? and that mss should be

something much bigger than mtu, which gives the performance improvement

shown in the numbers.
thanks,
jordi



From: David S. Miller

Subject: Re: TCP Segmentation Offloading (TSO)

Date: 	Mon, 02 Sep 2002 23:52:44 -0700 (PDT)
The performance improvement comes from the fact that the card

is given huge 64K packets, then the card (using the given ip/tcp

headers as a template) spits out 1500 byte mtu sized packets.
Less data DMA'd to the device per normal-mtu packet and less

per-packet data structure work by the cpu is where the improvement

comes from.



From: Jordi Ros

Subject: RE: TCP Segmentation Offloading (TSO)

Date: 	Tue, 3 Sep 2002 00:26:13 -0700
What i am wondering is how come we only get a few percentage improvement in

throughput. Theoretically, since 64KB/1.5KB ~= 40, we should get a

throughput improvement of 40 times. That would be the case of udp

transmiting in one direction, in the case of tcp transmiting in one

direction (which is the one you have implemented), since in average we have

(at most) 1 ack every 2 data packets, we should theoretically obtain a

throughput improvement of (40+20)/(1+20) = 3 (this comes from: without tso

we send 40 packets and receive 20 acks, this is, the cpu processes 60

packets; whereas with tso we send 1 packet and receive 20 acks, this is, the

cpu processes 21 packets).

However, we don't see in the numbers obtained neither an increase of

throughput of 300% nor a decrease in cpu utilization of such magnitude. Is

there any other bottleneck in the system that prevents us to see the 300%

improvement? (i am assuming the card can do tso at wire speed)
thank you,
jordi



From: David S. Miller

Subject: Re: TCP Segmentation Offloading (TSO)

Date: 	Tue, 03 Sep 2002 00:39:13 -0700 (PDT)
Because he's maxing out the physical medium already.
All the headers for each 1500 byte packet still have to hit the

physical wire, that isn't what is being eliminated.  It's just

what is going over the PCI bus to the card that is being made

smaller.



From: Scott Feldman

Subject: RE: TCP Segmentation Offloading (TSO)

Date: 	Tue, 3 Sep 2002 10:50:27 -0700 
Jordi Ros wrote:
> What i am wondering is how come we only get a few percentage

> improvement in throughput. Theoretically, since 64KB/1.5KB ~=

> 40, we should get a throughput improvement of 40 times. 
You're confusing number of packets with throughput.  Cut the wire, and you

can't tell the difference with or without TSO.  It's the same amount of data

on the wire.  As David pointed out, the savings comes in how much data is

DMA'ed across the bus and how much the CPU is unburdened by the segmentation

task.  A 64K TSO would be one pseudo header and the rest payload.  Without

TSO you would add ~40 more headers.  That's the savings across the bus.
> Is there any other bottleneck in the system that prevents

> us to see the 300% improvement? (i am assuming the card can

> do tso at wire speed)
My numbers are against PCI 64/66Mhz, so that's limiting.  You're not going

to get much more that 940Mbps at 1GbE unidirectional.  That's why all of the

savings at unidirectional Tx are in CPU reduction.
-scott

• Google archive of above thread