I'm not sure... I didn't investigate much further once I figured out it'd work 
fine for our application. This was sending data through a switch, so it could 
even have been due to the switch fabric. I didn't try a direct cable 
connection. But it might be there's overhead in our 10GbE core's CPU/fabric 
arbiter, for example. As I say, I didn't look too closely.

Jason


On 14 Mar 2014, at 11:55, Jack Hickish <[email protected]> wrote:

> Thanks Jason. That info is really helpful.
> 
> It sounds like if I boost my packets to ~1% header overhead, I
> shouldn't reckon on being able to throw away less than 5% of the
> bandwidth.
> 
> Out of interest, do you know what limit you're hitting at 96% data
> rate (I interpreted raw data rate to mean payload + headers(?))? -- is
> it something inherent to the tx/rx firmware or the switch, or is it
> ARP packets and the like floating round?
> 
> Thanks again,
> 
> Jack
> 
> On 14 March 2014 06:20, Jason Manley <[email protected]> wrote:
>> You can push them pretty hard. We're reliably running at an average of 
>> around 90% wire speed on all four links for MeerKAT. But be sure to 
>> factor-in the overheads of the core...
>> 
>> In our case:
>> L1 Ethernet HDR & FTR    44B
>> IPv4 header              20B
>> UDP header                8B
>> Application header       32B
>> Data payload           5120B
>>                      ======
>>                       5224B
>> 
>> We normally run each link at 8.733Gbps raw data rate, plus M&C data, ARP and 
>> all the rest. But I've done tests showing it working well up to about 96%. 
>> Beyond that I started losing lots of packets.
>> 
>> Jason Manley
>> CBF Manager
>> SKA-SA
>> 
>> Cell: +27 82 662 7726
>> Work: +27 21 506 7300
>> 
>> On 13 Mar 2014, at 20:08, Jack Hickish <[email protected]> wrote:
>> 
>>> Hi all,
>>> 
>>> I have an application that generates 40 Gb/s of data and I'd like to
>>> output it over 4 x 10GbE links (i.e. one roach2 mezzanine card).
>>> 
>>> I'm going to have to throw some data away to make room for packetising
>>> overhead, and right now I'm trying to work out how much, to see
>>> whether I should just go to an dual-mezzanine card setup. Is anyone
>>> with experience operating ROACH2 10GbE networks at near to peak
>>> capacity able to give me an idea of what throughput I might be able to
>>> get once I factor in overheads? For what it's worth, I'm using packet
>>> sizes of ~1kB in size in a network of 10 ROACH2s connected to a 48
>>> port switch.
>>> 
>>> Thanks,
>>> 
>>> Jack
>>> 
>> 


Reply via email to