Operating System - HP-UX
1833051 Members
2388 Online
110049 Solutions
New Discussion

HP-UX Auto Port Aggregation - support for 10GigE?

 
SOLVED
Go to solution
Jyotinath Ganguly
Occasional Contributor

HP-UX Auto Port Aggregation - support for 10GigE?

Today HP-UX Auto Port Aggregation s/w supports 10, 100 1000BaseT; is the roadmap for 10GigE suport known?
4 REPLIES 4
Luk Vandenbussche
Honored Contributor
Solution

Re: HP-UX Auto Port Aggregation - support for 10GigE?

Hi, it is supported, you see it in table 1-2 on the link below

http://docs.hp.com/en/J4240-90035/J4240-90035.pdf
rick jones
Honored Contributor

Re: HP-UX Auto Port Aggregation - support for 10GigE?

Are you looking for > 10 Gbit/s under a single IP, or are you looking for the HA aspects of APA?

BTW, what is your performance expectation with 10G?
there is no rest for the wicked yet the virtuous have no pillows
Jyotinath Ganguly
Occasional Contributor

Re: HP-UX Auto Port Aggregation - support for 10GigE?

Customer is looking for > 10 Gbit/s under a single IP, and > 1 Gbit/s sustained bandwidth.

Would IP over Infiniband be a good option for Enterprise (Insurance) applications using Integrity + HP-UX + Oracle Db + WebSphere stack?
rick jones
Honored Contributor

Re: HP-UX Auto Port Aggregation - support for 10GigE?

> 10 Gbit/s under a single IP is "interesting" what sort of system are we talking about?

You say Oracle and DB and such, that tends to mean small packets. In _very_ broad handwaving terms, 10Gig, just like 1 Gig and 100BT before do nothing to make transfer of data any easier on the host - it takes just as many compute cycles to send a packet over 10G as it does 1G or 100BT or 10BT.

Now, there are specifics of the various NICs which make things better - for example, the 1G NICs offer ChecKsum Offload (CKO) and "Large Send" (UX calls it VMTU) support, along with interrupt coalescing, which makes the specific NIC _implementation_ better than the 100BT. It also offers JumboFrame, but unless you can get some tcpdump/tusc traces showing sends > 1460 bytes, that may not really matter much in this case.

The 10G NIC (AB287A) continues/builds-upon that by offering multiple queues, so it can take advantage of multiple CPUs.

BTW, the AB287A is a PCI-X 1.0 card, which means 133 MHz on the PCI-X bus, which means that total bandwidth through the card, no matter how much CPU is behind it, is limited to 7.X Gbit/s, and there may be system-specific limits beyond that.

And just because I'm paranoid :) you do know that under APA, it is the _aggregate_ which is increased, not necessarily a single connection?

IP over IB - I'm not all that familiar with the IB cards. My understanding is they do not offer CKO. I would want to benchmark with say netperf configured with parms to simulate the expected workload.
there is no rest for the wicked yet the virtuous have no pillows