New on LowEndTalk? Please Register and read our Community Rules.
All new Registrations are manually reviewed and approved, so a short delay after registration may occur before your account becomes active.
All new Registrations are manually reviewed and approved, so a short delay after registration may occur before your account becomes active.
UDP packet loss inside same data center
leb_anonymous
Member
in Providers
I was running some UDP testing (using iperf3) between my VPS instances. I tried both linode and digitalocean.
I was shocked to find a LOT of packet loss!
For example,
iperf3, udp with 10mbps gave 15% packet loss from london to newyork.
Same iperf3 results gave 4% packet loss from london to london (just one single network hop!)
Almost equivalent results for linode and digital ocean. Imagine these are from the top providers!
Is this normal for VPS?! My dedicated servers never did that! (dedicated server germany-dallas, gave me 50mbps UDP traffic without a single packet loss!).
Is this a problem with iperf3?
Comments
No, that's not normal (unless they limit iperf testing). Have you contacted the support with your findings?
Can you see if the MTR between hosts will also show some packet loss?
For digital ocean, I wasn't able to reproduce the problem. Seems like a transient problem.
For linode, I have a loooong issue with many iperf/mtr logs (since more than 7 days).
Have you contacted the Datacentre with your findings?
What was their response?
"We sent you request to network team", "We forwarded it to upstream provider", "Send us iperf/mtr" and some other canned responses.
Maybe they just rate limit UDP?
Many datacentres rate limit UDP, what happens if you try the same tests with TCP? Do you get the same packet loss?
The key is to determine whether there is pathloss or just loss when using UDP. If its just UDP then there is likely to be some shaping or rate limiting taking place.
Private/Internal network or public network communicating internally?
I must admit I have been struck by UDP rate limiting with many providers, specifically on the receiving end (syslog receiver).
Internally on DO's private network I have seen packet loss / tcp re-sends occasionally. I suspect they have some QoS in place. The Elasticsearch nodes we run tend to push peaks of 200-300mbps over the network at times (its likely just a 1gbps port, shared).
It is public network communicating internally (iperf3 on public IP for VPS which is one hop far from me).
TCP works good (with few retries) in both cases (linode and DO). Is there any way to check for traffic shaping?
Updates about the UDP issue:
- Linode is really bad for UDP traffic, stay away from it. Customer support will deny the problem. Still using it for non-udp applications.
- Digital ocean works great for inter-Europe traffic. Sometimes there are hiccups in europe-usa links.
I found that dedicated servers are MUCH better in handling network-based applications. So I replaced multiple VPSs with small dedicated.
It has nothing to do with virtualisation (or lack of thereof). Don't mix the layers.