Aruba & ProVision-based
cancel
Showing results for 
Search instead for 
Did you mean: 

5400zl Series Switches: Slow Data Transfer between ZL1 and ZL2 Modules

 
Highlighted
BobLenz
Occasional Visitor

5400zl Series Switches: Slow Data Transfer between ZL1 and ZL2 Modules

Hello!

I have both 5406zl and 5412zl switches with J8702A 24 Port Gigabit Modules and J9546A 8p 10GBase-T v2 zl modules.  Copying Data between devices on the same module is fine: very fast. Also good between modules of the same module type.  I, however, have noticed a drop in data rates between the v2 and the v1 modules.  When copying from a server on a 10GBase-T line, It fluxuates greatly from 40-75 MBps.  If I plug into the J9546 directly with a 1 G device, I average 113 MBps on the transfer solid.  I get the same rates if I move the server down to 1G on the older 8702 modules.  The backplanes on these switches are not busy, not even close to what the 40 Gbps per module limit is for the switch.

This used to not be an issue with these switches.  I'm not sure what could have changed.  I have isolated the switches to see if there was an outside influence, but that didn't seem to change the outcome.

Both Switches were on K.15.18.0015.  Just upgraded them to K.16.02.0026 to see if anything changed, but ended up with the same results.

Any idea why this could be happening?  I verified that tcp-push-reserve was indeed off, which I had seen for issues on other sites.

Thanks for any insight!

Bob

4 REPLIES 4
parnassus
Honored Contributor

Re: 5400zl Series Switches: Slow Data Transfer between ZL1 and ZL2 Modules

That's interesting.

Out of curiosity, what is the connection pattern for 10Gbps ports (of the very same module)?

It could be something related to stepping through different packet buffer sizes (zl v1 versus zl v2 modules)?

Is a 10Gbps to 10Gbps data transfer test [*] (using ports of different channels to avoid oversubscription) executed between ports within the very same zl v2 module (or between two different J9546A modules within the very same chassis) showing the same behaviour?

[*] I presume you're performing tests using a tool like iperf in a reproducible and controlled way.
Highlighted
BobLenz
Occasional Visitor

Re: 5400zl Series Switches: Slow Data Transfer between ZL1 and ZL2 Modules

I first noticed this with file transfers between my primary file server and other devices on the network.  The file transfer rate would vary dramatically.  iperf seems to show a constant 112-113 MBps range, which is what I would expect from gigabit (going from 10G server to 1G Client).  However, that same client has very strange data rates when copying a file down averageing 50-60 MBps.  As mentioned, if I take that 1G client and put them on the same 10G module, the data rate runs in the 112-113 MBps range that I'd expect, same if I move them to a 1G module.  The server has plenty of capacity for reading/writing data.

When I run iPerf between different ports on the same 10G module (file server is plugged into port 5).  I have run tests from iperf "clients on ports 7 and 4 and average 2.5 Gbps max (which is also what I experience with file transfers).  I have a LACP trunk going from my 5406 in the server rack to a 5412 "core" switch that consists of ports 1-3 on the 10 GB Module on each end.  When I run a iperf from the 5406 to the 5412's 10G module (port 5 on the server side(5406) and port 8 on the core side (5412), through the LACP trunk, I get about 1.6 Gbps max.  It would seem there is a bottleneck issue with the 10G modules, but there are no log errors on either switch and Error counters are clear.

Thanks for your reply!  Let me know if there is any further information I can provide.

Highlighted
BobLenz
Occasional Visitor

Re: 5400zl Series Switches: Slow Data Transfer between ZL1 and ZL2 Modules

Ok... iperf noob here.  I realized that it need to have the -P paramater with more instances (used 20) and got ~9 Gbps between port 4 and 5 on the same module.  Over the LACP trunk, though, between the 2 switches, I only got ~3 Gbps with the same -P 20 test.  This was on a VM for the client with a host connected @ 10G as I don't have a physical server "connected" to that port, so I'll try this with another physical device in the morning to see what results I get.

Highlighted
parnassus
Honored Contributor

Re: 5400zl Series Switches: Slow Data Transfer between ZL1 and ZL2 Modules

OK, testing with iperf is a good way to have reproducible results (if parameters are consistents across all tests).

I would suggest you this article interesting:

https://support.hpe.com/hpsc/doc/public/display?docId=mmr_sf-EN_US000005378

it's about oversubscriprions, packet buffers and performances of zl v1 and (and versus) zl v2 modules.

Pay attention to channels and connection patterns of 10Gbps interfaces to avoid potential oversubscription scenarios.

With regard to Port Trunking...you have to consider that host-to-host data flow (even if those hosts are connected to different switches uplinked together via LACP Port Trunk) typically shows a stream that, from the point of view of LACP hashing, will traverse a 2 ports LACP ports trunk always using one of both links (it's biased)...so typically this stream will benefit from resilience, less (if none at all) of enhanced throughput provided by two aggregated links...a way to see the Port Trunk really used at its limits is to create various data streams (from and to differents IPs and Ports)...that way it's much more probable that both links will be used concurrently...that's to say that I don't expect automatically a 2x speed when I see a 2 links aggregated uplink if generated data stream is provided by one host and the destination is just another host on the other side of the uplink.