Hi Jeff,

On 05/09/2014 06:11 PM, Shaw, Jeffrey B wrote:
> I agree, we should wait for comments then test the performance when the patches have settled.

Here are some performance numbers I've measured with the TSO
patches. The test platform is:

+-----------+           +-----------+
|           |           |           |
| traffic   |-----------| dpdk      |
| generator |-----------| testpmd   |
|           |-----------|           |
|           |-----------|           |
|           |           |           |
+-----------+           +-----------+

- 4 ixgbe ports
- sandy bridge at 2.7 Ghz

I've only included numbers for pkt_size=64. Other packet sizes
do not bring more information in this case.

I have 4 test cases:

- testpmd in iofwd mode with normal tx/rx function
- testpmd in iofwd mode with simple tx/rx function (txqflags=0xf01)
- testpmd in macfwd mode with normal tx/rx function
- testpmd in macfwd mode with simple tx/rx function (txqflags=0xf01)

I tested this for 1c1t, 1c2t, 2c2t, 2c4t, 4c8t on the following version:

- dpdk.org head
- dpdk.org + tso patchs until 6/11 (included): it includes all mbuf
   reworks (data_offset instead of data, remove ctrl mbuf, use 48 bits
   physical address)
- dpdk.org + all tso series

The conclusion of the tests is:

Patches up to 6/11 do not bring any performance regression. On the
other hand, the full TSO patch series introduces a small performance
regression (usually corresponding to ~5 cycles per packet). This is
probably due to additional tests related to TSO done in driver.
I suppose that this performance loss is acceptable if we consider
that TSO will bring a huge performance enhancement for many real use
cases.

By the way, I found lower numbers in macfwd mode + simple rx/tx
with current version (without my patches) with 1c2t. It seems
reproduceable.

I'll soon provide a v2 that will include:

- the split of patch 6/11 (cosmetics vs functional)
- the split of patch 11/11 (ixgbe vs generic changes)
- new checksum flags PKT_RX_L4_CKSUM_GOOD and PKT_RX_IP_CKSUM_GOOD
   proposed by Stephen
- modifications of external PMDs (memnic, virtio, vmxnet3)

Regards,
Olivier