Session: DCIM-B344, Network Turning for Specific Workloads. This was a great session, with a ton of Q&A during and after the main presentation was over. I’d highly encourage you to watch the full video on Channel 9 when it is uploaded to get all of the goodies. The session notes below are a small fraction of the gold nuggets that were discussed in the session. Confused about VMQ, RSS, vRSS, SMB multi-channel performance, virtual switches, NIC teaming and when to use what feature? Be confused no more after watching the video.
Terminology:
- Socket is a NUMA node, and within the node is a core. On the core you have logical processors (with hyper-threading), on which you have virtual processors for VMs.
Scenarios
Problem 1: Enterprise physical web server and file server. Large volume of incoming packets, but one core is highly utilized.
Solution: Enable RSS on the server. RSS is for physical servers only. NIC spreads the network traffic by TCP/UDP flows across different cores to enhance performance and balance processor utilization.
Problem 2: A VM is deployed and the incoming packet processing is saturating a limited set of cores.
Solution: Virtual machine queue. VMQ spreads traffic per vNIC. RSS is disabled on the pNIC when a virtual switch is defined. A single core is bottlenecked at 4-5 Gbps of traffic, depending on processor speed. VMQ is enabled by default, so no manual configuration is needed. Number of queues depends on the physical NIC properties. New NICs have more queues (64+ not uncommon).
Problem 3: A VM has a large number of incoming packets, such as a web server. The workload is limited to using one vCPU. This is only for VMs with >3 Gbps of traffic. Less traffic can be serviced by a single core without any additional configuration.
Solution: vRSS can be used on WS2012 R2 VMs. This spreads traffic across multiple vCPUs. Flows are moved if a CPU has 90% or higher utilization. MS states they have seen line rate up to 40Gbps to a VM using vRSS with a 40 Gbps NIC. vRSS must be manually enabled inside of the VM.
Problem 4: A highly latency sensitive application, such as high speed financial trading.
Solution: Use SR-IOV. Bypasses the virtual switch, and directly connects the VM to the hardware NIC. Only for use with trusted VMs, since switch security is bypassed. Rarely used, but available for these very limited cases.
NIC Teaming
Windows Server 2012 R2 has a new dynamic NIC teaming mode. Continuously monitors traffic distribution. Actively adjusts traffic based on observed load. Download the Windows Server 2012 R2 NIC teaming guide here.
