cross-posted from: https://infosec.pub/post/306795
I am interested in your ways to identify a bottleneck within a network.
In my case, I’ve got 2 locations, one in UK, one in Germany. Hardware is Fortigates for FW/routing and switches are Cisco/HPE. Locations are connected through an Ipsec VPN over the internet and all internet connections have at least a bandwidth of 100 Mbps.
The problem occurs as soon as one client in UK tries to download data via SSH from a server in Germany. The max download speed is 10 Mbps and for the duration of the download the whole location in UK has problems accessing resources through the VPN in Germany (Citrix, Exchange, Sharepoint, etc).
I’ve changed some information for privacy reasons but I’d be interested in your first steps on how to tackle such a problem. Do you have some kind of runbook that you follow? What are common errors that your encounter? (independently from my case too, just in general)
EDIT: Current list
- packet capture on client and server to check for packet loss, latency, etc. - if packets dropped, check intermediate devices
- check utilization of intermediate devices (CPU, RAM, etc)
- check throughput with different tools (ipfer3, nc, etc) and protocols (TCP, UDP, etc) and compare
- check if traffic shaper/ QoS are in place
- check ports intermediate devices for port speed mismatch
- MTU/MSS mismatch
- is the internet connection affected too, or just traffic through the VPN
- Ipsec configuration
- turn off security function of FW temporary and check if it is still reproducible
- traceroute from A to B, any latency spikes?
- check RTT, RWND, MSS/MTU, TTL via pcap, on the transferring client itself and reference client, without and while an active data transfer
Prob not related but noteworthy:
- check I/O of server and client
I’ll keep this list updated and appreciate further tips.
Update I had to postpone the session and will do the stress test on Monday or Tuesday evening. I’ll update you as soon as I have the results.
Update2 So, I’ll try to keep it short.
First iperf3 over TCP run (UK < DE) with same FW rules let me reproduce the problem. Max speed 10 Mbps, and DE < UK even slower, down to 1-2 Mbps. Pattern of the test implies an unreliable connection (short up to 30 Mbts, then 0, and so on). Traceroute shows same hops in both directions, no latency spikes, all good.
BUT ICMP and iperf3 over UDP runs show a packet loss of min 10% and up to 30% in both directions! Multiple speed tests to endpoints over the internet (UK>Internet) showed a download of 80 Mbts andupload of like 30 Mbts, which indicates a problem with the IPSec tunnel.
Some smaller things we’ve tried without any positive effect:
- routing changes
- disabling all security features for affected rule set
- removed traffic shaper
- Port speed/duplex negotiations are looking good
- and some other things that I already forgot
Things we prepared:
- We have opened some tickets at our ISPs to let them check it on their site > waiting for response
- Set up smokeping to ping all provider/public/gw/ipsec endpoinrts/host IPs and see where packets could be dropped (server located in DE)
- Planned a new session with an Fortigate expert to look in-depth into the IPSec configuration.
Need to do:
- look through all packet captures (takes some time)
- MSS/MTU missmatches / DF flags
- further iperf3 tests with smaller/larger packet
- double check ipsec configuration
- QoS on Switches
I wish I had more time. I’ll keep you updated
Update3 Most likely the last big update.
So, the actual infrastructure is a little bit more complex than I’ve described in this post, so nobody could have suggested tips for this case.
We think that we have found the problem, but we couldn’t implement the fix yet since it requires some downtime, and I was on a business trip. We’ve got multiple locations in the UK that are connected to a third party (MLPS) where their internet breakout points are too. We’ve now got multiple IPSec tunnels that terminate on the same FW in Germany. The problem is that the third-party FW uses the same IP AND port for all IPSec tunnels too, which most likely causes all the issues. In short: only use one tunnel or change the GW on the German side.
Don’t ask me why, please! - It is a cluster fuck, and the goal is to fix it in the future. One site had a large flat /16 network not long ago.
I might share a final update when we get the fix implemented.
Care to tell me those numbers?
Honestly, that sounds like TCP bandwidth-delay product.
50 ms with a 65k byte RWND is just around 10Mbps.
See for yourself with your numbers:
https://wintelguy.com/wanperf.pl
Just saw this part:
Some VPN solutions downgrade the MSS of all VPNs to the lowest common denominator for things like MTU/MSS. I guess that can make sense in a full-mesh, but whatever.
Take a packet capture of another client while the problem one connects, you’ll likely see something.
Decrypted traffic is usually easier to analyze.
Ohhh and you say that’s when they connect through SSH? Check that he’s not tcp forwarding all traffic through his SSH connection somehow.
Getting a pcap of another client could bring some insight, yeah.
SSH is used for the data transfer. Without knowing it at this moment, I’d assume scp or rsync. You mean whether all their internet traffic is routed through the active SSH session?
I mean that in an SSH connection you can configure it to bind local/remote ports of local/remote IPs.
The user might have unknowingly or maliciously configured their stuff to either:
Unlikely, because they couldn’t bind a port that is already in use on the server. Still, that could technically happen if there’s a misconfigured load balancer, maybe from an old config that was never removed, that has that server as a member and just declares it down/up when that user starts listening on that port.
That last one is far-fetched.
I’d start with cpu/mem, mtu/mss, etc.
I tend to have a bit of a bias towards absolutely far-fetched things because I’m basically the last line of support where I work. This means
allmost of the “normal” problems get filtered out before they get to me, which leaves me with the stuff that’s bananas.I’ll keep that in mind
Ping - Update 2 Your numbers are are still missing since I havent had time to look into the pcaps yet. I hope I can get it done by the end of the week, but we are a little bit wiser.
I haven’t had the chance to get a pcap yet. As soon as I get my fingers on the test clients, I’ll check them and additionally do testing with TCP and UDP transfers. I’ll let you know.
Just to clarify: this would be the limit for a single TCP connection and yes, could be the limit for this one download. This would not explain, why the rest of the location is affected if theoretically 90% of the bandwidth is still available, no? - Please correct me if I am wrong here.
Yea my bad, I missed the “whole location” bit on first read. This would be the limit for that TCP session.
Still, I’d compare MTU, MSS, RTT packet loss, RWIND, etc. everything that is a component of the actual bandwidth. Whatever happens, some of these things change when he connects.
I’ve had VPN solutions that downgrade the MTU for everyone when someone with a shit MTU connects to it.
Another thing that came to mind, since you were talking about an SSH connection, is if that user is somehow routing trafic through a tcp forward inside his tunnel when connecting. Stupid test maybe, but I’d compare the before/after TTL of packets in a flow that is known to be affected as well as a traceroute (assuming the client can even run one, because every business seems to like breaking icmp)
Will compare it as soon as I get my hands on the machine.
And yeah, we do tend to block ICMP over here too.