Netscaler VPX 1000 - Azure - Slowness getting through Netscaler.
Article ID: CTX231986
Updated On:
Description
Only one single interface 0/1.
TX Overflows and interface congestion counters got increase aggressively
19 7274 164576686 64911 8923 nic_err_tx_overflow interface(0/1) Wed Jan 10 19:23:12 2018
16 17032 164402447 83975 8718 nic_err_tx_overflow interface(0/1) Wed Jan 10 19:22:50 2018
16 17032 164402447 83975 8718 nic_err_tx_overflow interface(0/1) Wed Jan 10 19:22:50 2018
18 7049 164511775 108174 15346 nic_err_tx_overflow interface(0/1) Wed Jan 10 19:23:04 2018
18 7049 164511775 108174 15346 nic_err_tx_overflow interface(0/1) Wed Jan 10 19:23:04 2018
3 0 20407 493 69 nic_err_congested_pkts_dropped interface(0/1) Thu Jan 11 01:10:26 2018
3 0 20407 493 69 nic_err_congested_pkts_dropped interface(0/1) Thu Jan 11 01:10:26 2018
17 0 18542 539 50 nic_err_congested_pkts_dropped interface(0/1) Wed Jan 10 22:30:37 2018
17 0 18542 539 50 nic_err_congested_pkts_dropped interface(0/1) Wed Jan 10 22:30:37 2018
Resolution
With 12.0 builds, we have changed default yield behavior for PE vCPUs. vCPU will not yield to hypervisor, even though if there is less/moderate traffic in 12.0 build, which was not the case for 11.1 builds. That’s the reason, VPX vCPU is always 100% on hypervisor. However, vCPU is allocated to management core might not be 100%.
NetScaler yields PE vCPUs to hypervisor in sparse/moderate traffic cases. Since we have observed Tx overflow/congestion, it’s somewhat related to scheduling, we thought not yielding vCPU helps in improving the situation.
- set ns vpxparam -cpuyield NO
Upgrade to 12.0.53.X+
Issue/Introduction
