Score:0

What LAN issue is characterized by very slow upload speed until an upload to a different target is started on which both get normal?

in flag

When I upload something from a Windows server (qemu kvm) to a new Linux server (bare metal) in the same network the speed is very slow (about 1/100 of what should be possible for the 1GBit/s uplink). Uploads to all other machines (including other Linux servers) in the network work at full speed. And as soon as I start such an upload to another machine while an upload to the problematic Linux server is still running, both uploads get fast (so the before slow upload increases to about 50% of the speed of the uplink while the other starts and stays at that as well). Once the 'other upload' finishes the former upload to the new problematic server drops to its very slow speed.

This seems to be the case for all traffic (SSH, HTTP, SMB) while no other machine in the network has the problem. So every other machine in the network uploads to the new server without issues at full speed. It even seems that the Linux bare metal host also has no issues.

Between both servers are two Netgear 1/10GBit/s switches but no VLANs or any other special configuration. I tried some typical KVM host/guest workarounds (tx/rx offload, lso, different virtual adapter, ...) but without any change at all. Looking at tcpdumps on source, target and host I also don't see anything which seems off. So no package loss or other problems I could identify (though I'm no expert here).

So now before anything else and since I've never seen something like that, my main question is at what kind of problem am I even looking here?

Score:0
cn flag

My first guess would be there's something wonky with ethernet auto-negotiation between the Windows VM and the baremetal Linux box that causes the Windows VM's "port" to negotiate to some lower level, like 10M instead of 100M or 1G. When the Windows VM uploads to a different server, the auto-negotiation problem isn't there (or rather, is overruled so long as the connection to the other server is active) and the port uses 1G.

in flag
Sounds like a logical conclusion. How would I debug it, though? I just checked `Get-NetAdapter | SELECT name, LinkSpeed, fullduplex` on the Windows VM and `ethtool eth0` on the host during the slow transfer was running and all ports here show the expected speed.
cn flag
Hmm. The first thing I'd do is run `Get-NetAdapter` and `ethtool` both during the slow transfer itself, and then during the dual transfer where things speed up. From there, I'd compare the output for each between "slow" and "dual" and search for differences. Also, is the windows VM you mentioned the only windows host? I.e., are all other hosts on the network some flavor of GNU/Linux?
mangohost

Post an answer

Most people don’t grasp that asking a lot of questions unlocks learning and improves interpersonal bonding. In Alison’s studies, for example, though people could accurately recall how many questions had been asked in their conversations, they didn’t intuit the link between questions and liking. Across four studies, in which participants were engaged in conversations themselves or read transcripts of others’ conversations, people tended not to realize that question asking would influence—or had influenced—the level of amity between the conversationalists.