Bug #7426
closed
Added by Chris Macmahon over 7 years ago.
Updated about 7 years ago.
Affected Architecture:
SG-1000
Description
When doing an iperf test outside of pfsense there is a strange packet loss at the start of the test.
UDP packets drop
TCP packets go through.
Was able to confirm it's isolated to an SG1000, on 2.4 SG-2220 no issues were present in this test, but we were able to reproduce on SG1000.
See ticket 15982
Files
I made some tests with simple DNS answer/reply and noticed problem with states overload (250 clients and 250 servers which is 62500 states in total vs 49000 by default) and UDP drops. Just increased states to 500000 and there was not UDP drops until huge UDP load (10000 answer/reply in second)
- Target version changed from 2.4.0 to 2.4.1
The reason of UDP drop is packet processing slowdown which happens on ARM devices (1k, 3100). I observed ~2-7% for different packet size and one does not change if bandwidth is changed (provided that packet size is not changed). It does not affect data transmission in whole, but UDP tests show drop because UDP traffic generator knows nothing about delay.
Picture below demonstrates this, blue line is number of RX packets one is the same every time ~24.5kpps which is maximum performance for sg1k on 64byte frames.
And no drops for DNS request/answer (~5000pps)
---------------
port : 0
------------
opackets : 112499
obytes : 8662423
ipackets : 112502
ibytes : 10462483
Tx : 1.31 Mbps
port : 1
------------
opackets : 112499
obytes : 10462407
ipackets : 112506
ibytes : 8662871
An important note
The issue can be found on:
- sg1k for ordinary data transmission
- 3100 for VPN data transmission when CPU is highly loaded
- hardly ever on intel devices, I saw it several times for VPN data transmission with hight CPU load and huge number of packets (usually small/random)
- Target version changed from 2.4.1 to 2.4.2
- Project changed from pfSense Packages to pfSense
- Target version changed from 2.4.2 to 2.4.3
- Status changed from New to Feedback
- % Done changed from 0 to 100
Fixed. Add nmbclusters set to 500000 on SG-1000 and 1000000 on SG-3100.
Tested iperf3 in UDP mode between SG-1000 and SG-2440 (and SG-3100), could not reproduce the bad behavior. Saw a problem on one device, but it appears the issue there was the client's nic (not the SG-1000 itself). Tested with pfSense-netgate-uFW-recover-2.4.2-DEVELOPMENT-armv6-20171104-1805
- Status changed from Feedback to Resolved
- Target version changed from 2.4.3 to 2.4.2
Also available in: Atom
PDF
Updated by Constantine Kormashev 
Updated by 
Updated by 
Updated by Anonymous