Will APA Trunking Result in increased Packets/Second Capacity?

[Alzhy]

We're thinking of Trunking 2 Gigabit Cards in the hope of increasing bandwidth and network (packets/second) throughput to one partricularly busy HP-UX 11.11 server that serves over 2 Dozen Gigabit connected Windows servers.

We already empoy TSO (TCP Offloading) but it seems the Apps still clamour for more bandwidth. Will trunking solve this problem? Or will I be better off segmenting LAN or multi-homing my HP-UX Server -i.e. I establish more APA (LAN Monitor) groups - with each APA Interface pair serving solely a group of NT Servers that we deem can be supported by each APA pair?

[TwoProc]

We APA truncd over here and it makes a difference. You'll notice that for a single process, however, you won't gain. It look like it's not until you overflow the capacity of a single port with secondary and tertiary (for example) processes until the next port starts doing anything. But, for an Oracle setup with lots of connections, making the most of it won't be a problem. You can put at most 4 ports together into a single switch. But, the throughput of 4 2Gs should provide you with lots of bandwidth without restructuring switch flows.

[Kevin Wright]

Trunking will certainly increase your bandwidth yes, since you already apparently have APA SW, you should configure a trunk group and check your usage.

[rick jones]

TSO is "simply" TCP _Segmentation_ Offloading, not to be confused with TOE or TCP Offload Engine, which is the term for "full" TCP protocol offload.

Anyhow, unless your applications are making send() calls of > 1460 bytes at a time, you will not get much out of TSO. So, you might want to tusc trace your applications to see what size send calls they are making.

You should also verify that "TOPS" is enabled on the system, particularly if you see one or more of your CPUs close to saturation with interrupt processing.

Just trunking two NICs is probably fine, but if you want to trunk more than two together, you want the "NO_SYNC" patches, and I cannot recall if those are available back on 11.11. If they are not, then you will want to start to segment your network and have multiple "interfaces" (from the standpoint of the transport) on the system, each (preferably) configured into a separate IP subnet.

[Muralidhar Venkat]

"NO_SYNC" patches are available on 11.11. See APA performance paper @
http://docs.hp.com/en/7662/new-apa-white-paper.pdf

[Jim Keeble]

The way the data is distributed has nothing to do with filling the capacity of the first link. Rather it depends on the load balancing algorithm you select when configuring the aggregate.

You mentioned Lan Monitor groups, which do no load balancing at all. They provide redundancy, switching all the traffic to the standby interface when the primary fails.

If you are using load balancing aggregates, the default load balancing algorithm is based on the destination mac address. If you are sending only to one mac address at a time, all the traffic goes out one physical link. And, if the Windows hosts are on a different subnet, then all that traffic goes out of a single port because all the traffic goes to one mac address, the router's.

The best load balancing algorithm is usually LB_PORT, which establishes the "data flows" for distribution based on the sum of the source and destination TCP or UDP port numbers in each packet. In this way, traffic can be distributed even if all the traffic at some point in time is flowing all to the same remote host.

If the traffic traverses other trunks along the way (through routers or between switches or even at the remote system) the load balancing over each trunk is determined by the sending link partner. So, the HP does it's load balancing when giving packets to the switch, but the switch does the load balancing when forwarding the packets through the next trunk. The switch may also support load balancing based on the tcp and udp port numbers, some do, some don't.

Hope this helps.

[Steven E. Protter]

Shalom,

I think you will be disappointed in the results.

A normal APA setup with a pair of GB cards should get you twice the bandwidth of a single card without APA.

There is some overhead associated with APA and the type of requests, condition of the network, collision domain size and many other factors will hurt your bandwidth.

We had a similar issue with bandwith to and from our central storage NFS system, which is an EMC. We solved it mostly with networking.

Some systems were granted direct fiber access to the fabric network, negating the need to use copper network bandwidth.

A special network for only NFS was configured with a much smaller collision domain. At the very least our NFS packets were not bouncing off our Windows SMB packets.

The presence of Windows boxes in this mix leads me to believe the problem is not on your HP box. Its probably network configuration and what microsoft presents to the world as code.

I've noticed in a networks I maintain for fun that every windows box in the place was bouncing smb packets of my web servers. My solution was to physically disconnect the web servers from that network. Still every windows box in the place is filling my my firewall logs with SMB packets. What a waste.

SEP

SEP

[TwoProc]

SEP,

I'm confused, you say that he'd be dissapointed, but then you say that he should get nearly twice the throughput... which should make him more satisfied with the result.

I know it's not my thread, but can you clarify? Did you mean that even though his potential throughput would double his problem with the Windows Servers is the real issue there? I wouldn't mind learning a bit more from you on this issue, as well.

Thanks,

John