• Status: Solved
  • Priority: Medium
  • Security: Public
  • Views: 1217
  • Last Modified:

"Total Output Drops"

Could "Total Output Drops" cause resynchronizing of TCP communications?  For example with my network monitoring tool, I take an hour snapshot of replication data over the WAN in 1 minute samplings.  It's mountains and valleys.  I noticed on the outgoing interface that it is taking output drops, about 12,000 an hour.  Could it be that when TCP is negotiating the Mb/s climbs then with the output drops so does the speed and then starts the negotiations all over again?  Or is there another reason why the data doesn't just level off.  I was thinking Global Synchronization maybe?
2 Solutions
WissamSenior Network EngineerCommented:
if there is a larger burst of packets that come in a short span of time, and there is not enough space left in the {{output buffer}} to store all these packets at once, the port will store as many as possible till it fills the output buffer and drops the rest.
Once the space is freed on the buffer by transmitting some packets over the link, the other packets that arrive at a later point of time will be buffered again and transmission continues. I’ll use an example here for a better understanding, if the buffer size is 2MB, and 1MB is already used up, if the output interface receives 1.5MB of packets at once, then only 1MB of packets are buffered in the free space left and the rest (0.5MB) of the traffic is dropped. Hence these output drops on the interface increments. Hope this helps.
What is your WAN speed?  If it's considerably slower than the LAN speed (and total throughput expected based on latency to the remote end), then yes, you are exactly right and you're seeing TCP doing it's thing with peaks/valleys.

For example, if you have GigE-connected servers but only a 10Mbit/s link with 10ms latency, you'll consistently cycle into TCP slow-start.

A few questions:

1) What program are you using to copy that data?  If you have the ability to set a data rate that matches up with available WAN bandwidth, you should expect to see more consistent throughput / utilization.

2) What make/model is your WAN router, and what are the WAN/LAN interfaces (type/speed/count).  Do you know what type of queuing is used (tail drop, WRED)?
laz01Author Commented:
WAN speed is OC-12. The data is NetApp snap mirror backup.  The servers are etherchanneled with 4-5 Gig ports to server switch then a couple switches in between connected via 10gig then the OC-12.
All that's going on with flowcontrol is FIFO
Yep, then you hit the nail on the head. . . you are sending Gig's of data on a 622Mbit/s link.  Using FIFI you will have "tail drop" which as you mentioned increases the chance of global synchronization, especially so if all devices are using the same OS (and therefore same TCP settings).

At that rate, you will most definitely exhaust any packet buffers on the router which is intended to help smooth out TCP congestion control.

What make/model is your WAN router - and does it support advanced queue management?  See http://en.wikipedia.org/wiki/Random_early_detection as a good starting guide.
This question has been classified as abandoned and is closed as part of the Cleanup Program. See the recommendation for more details.

Featured Post

[Webinar] Cloud and Mobile-First Strategy

Maybe you’ve fully adopted the cloud since the beginning. Or maybe you started with on-prem resources but are pursuing a “cloud and mobile first” strategy. Getting to that end state has its challenges. Discover how to build out a 100% cloud and mobile IT strategy in this webinar.

Tackle projects and never again get stuck behind a technical roadblock.
Join Now