ndt-users - Re: Performance measurement boxes evaluation
Subject: ndt-users list created
List archive
- From: Alan Whinery <>
- To: , Performance Node Users <>, , ,
- Subject: Re: Performance measurement boxes evaluation
- Date: Wed, 07 Dec 2011 10:59:39 -1000
On 12/7/2011 10:15 AM,
wrote:
> I appoligize for the cross-post and multiple copies. If this
> information is documented please point me to it, as I have not
> stumbled across it.
>
> Does anyone know of or has anyone done evaluations of platforms
> for performance measurement boxes for the following categories?
>
> 1GE - End device to ping and test against
> - USB dongle or minimal device.
> - Place many of these in network for test points.
For any sort of measurement or performance test, avoid USB Ethernet like
the plague that it is. For one thing, stochastic Ethernet frames
traveling in stochastic USB packets seems to be a low-ball maneuver, for
throughput or for delay, and for another, the Linux drivers can
sometimes suck very much.
Some laptops do pretty well with built-in NICs, but YMMV unless a
specific model/chipset has been tested.
> 1GE - NDT, iPerf, I2 E2E Test suite, etc.
> - with 2 - 1GE ports and horse power to genereate 1GE sustained
There's a three-pronged inquiry to apply -- if you get on-board, PCIe or
DMI attached NICs, AND they use chipsets with good operational records,
AND Linux drivers with good operational records, then relatively small
(by current standards) Bogomips are required to push 1 GE. I bought:
http://www.supermicro.com/products/system/1U/5015/SYS-5015A-PHF.cfm?typ=E
for pS-NPT nodes of the "delay" flavor, but they're capable of keeping 1
GE busy on an extended iperf test without breaking a sweat. Sadly, they
only have x4 PCIe, so I can't try 10 GbE.
In my experience, one should avoid 1 GE devices that use linux modules:
sk98lin (superceding "sky2" seems to be much better)
forcedeth ("deth" should be pronounced "death")
> 10GE - NDT, iPerf, I2 E2E Test suite, etc.
> - with 1GE and 10GE port and horse power to genereate 10GE
> sustained
But then again, filling the pipe sometimes isn't the metric you'll get.
I have less experience with 10GE, but it reminds me that a 1GE machine
that is capable of only about 800 Mbps is pretty much the same as 950
Mbps WRT validating a path, and with 10GE, you're going to reach the
point where you're pretty impressed with 8 Gbps, or 6 Gbps on some
paths. All the 10 GbE NICs I have seen in real life are x8 PCIe, which
should work on PCIe 1.0, which is getting to be a clearance-table item.
It seems as though 10 GbE, with less bus/CPU/memory-bw headroom
(perhaps) is more finicky. More than once, I have upgraded a device from
1G to 10 G and had a performance drop on iperf tests, until I fiddled
with configurations.
If you haven't looked, fasterdata.es.net is a good place to find some
pertinent stuff, and perusing the hardware from the Internet Land Speed
Record competitions is illuminating, although getting to be out of date.
It could be useful to start building a page of what works, and what
doesn't. Maybe we should take that on.
Maybe we should pick a list, if you want to continue, interested people
who get left behind can contact directly.
> Thanks,
> -Roy Hockett
>
> Network Architect,
> ITS Communications Systems and Data Centers
> University of Michigan
> Tel: (734) 763-7325
> Fax: (734) 615-1727
> email:
>
>
- Performance measurement boxes evaluation, royboy, 12/07/2011
- Re: Performance measurement boxes evaluation, Alan Whinery, 12/07/2011
Archive powered by MHonArc 2.6.16.