Forum Discussion
DFresh_4130_150
Nimbostratus
NimbostratusSlow uploads through BIG-IP
We have a web app running behind an F5 BIG-IP running version 11.6.0. The app is responsible for storing large binary files so it regularly gets large uploads and downloads. During initial setup a while ago I was able to get the download speeds increased to around 30Mbps which was acceptable. We did this by updating the TCP profiles and found the Memory Management settings directly correlated to significant improvements. Below is what we're using for both client and server TCP profiles with the tcp-lan-optimized and tcp-wan-optimized profiles being the parent. We also have no iRules applied to this VIP.
Proxy Buffer High: 2516544
Proxy Buffer Low: 1258272
Receive Window: 1179630
Send Buffer: 1179630
Nagle: Disabled
The receive window and send buffer are the ones that have direct correlation to the increases in download speeds. I've been trying to increase our upload speeds, but can't seem to find any setting that makes a difference. If I upload directly from the node the app runs on itself I'm getting 30M+ upload speeds, but my desktop running curl shows around 975k on average. Output from curl on upload and download is below. Any suggestions on what to try next or how to troubleshoot?
C:\Temp>curl -u user -T file.mp4 -X PUT https://app.example.com/abc/test-generic/ > nul
Enter host password for user 'user':
% Total % Received % Xferd Average Speed Time Time Time Current
Dload Upload Total Spent Left Speed
5 616M 0 0 5 34.7M 0 975k 0:10:46 0:00:36 0:10:10 1025k^C
C:\Temp>curl -O https://app.example.com/abc/test-generic/file.rpm > nul
% Total % Received % Xferd Average Speed Time Time Time Current
Dload Upload Total Spent Left Speed
100 117M 100 117M 0 0 22.1M 0 0:00:05 0:00:05 --:--:-- 26.1M
Leonardo_SouzaCirrocumulus
First thing is to understand at which layer is the delay. Is that because of interfaces errors, TCP settings, HTTP profile, etc...
Interfaces errors, if easy to check in the interface status.
tmsh show net interfaceFor TCP, you need to get a tcpdump in the F5, and look for transfer pauses, zero windows size, etc..
For HTTP profile, is easy to rule out, just remove the profile from the virtual server and any client ssl or serverssl. You could also change the virtual server to forward IP, in case you are using the default standard type.
After you narrow down the problem, is easy to give a suggestion about what could be.
For tcpdump, check this solution:
John_Heyer_1508Cirrostratus
It may the that the proxy buffers and receive windows are actually too big. I'd recommend setting the proxy buffers back to 131072, but do increase the receive window to 256 or 512 KB. Send buffer of 1 MB is probably fine.
For the server profile, you can generally use tcp-lan-optimized, unless there is high latency, loss, or congestion expected between the BigIP and the backend servers.
Here's my blog post. I've been able to achieve 100 Mbps uploads and 250 Mbps downloads.
Noushad_VelladaI think F5 is buffering entire HTTP POST request before it starts transferring data to the pool member. So obviously this, in turn, increases the delay
- Leonardo_Souza
If you have an HTTP profile and a standard virtual server, it is normal.
Check this solution for more details:https://support.f5.com/csp/article/K8082
Noushad_Vellad1Altostratus
I think F5 is buffering entire HTTP POST request before it starts transferring data to the pool member. So obviously this, in turn, increases the delay
- Leonardo_Souza
If you have an HTTP profile and a standard virtual server, it is normal.
Check this solution for more details:https://support.f5.com/csp/article/K8082