WSV329: Architecting Cloud Infrastructure Using Windows Server 2012

Yigel Edery, Principal Program Manager, Microsoft
Joshua Adams, Senior Program Manager, Microsoft

This was a high level overview of various Hyper-V enhancements that make Windows Server 2012, cloud ready, according to Microsoft. It walked the audience through some architectural considerations when deciding on what technologies you should look at.

Windows Server 2012 is Cloud Optimized: Multi-tenant clouds, high scale and low cost datacenters, manageable and extensible
Summary

Dynamic and multi-tenant: Network virtualization, QoS, performance metrics, Live and Storage migrations
High scale and low cost compute: Larger hosts, large VMs, large clusters
High scale and low cost network: DCB, SR-IOV, RDMA, NIC teaming
High scale and low cost storage: Hyper-V over SMB, ODX, storage spaces, thin provisioning, synthetic fibre channel
Manageable and extensible: PowerShell, Hyper-V extensible switch

Datacenter Reference Architecture
Primary considerations: Workloads, Networking, Storage, Resiliency
Understanding Workloads

Cloud-aware stateless apps or legacy/stateful apps?
Workload performance requirements – 2 socket servers usually offer best ROI, apps networking patterns and the need for SR-IOV, mixing different servers to serve different workloads
Are workloads trusted? Level of isolation between workloads, QoS policies

Networking

Primary considerations: Isolation of traffic flows at physical and virtual level, type of infrastructure, NIC offloads
Typical Hyper-V server traffic flows

VM traffic
Cluster traffic
Storage traffic/CSV
Live Migrations
Management

How many NICs do I really need on each server?

WS2012: Run everything through the virtual switch, one physical network
Use Port ACLs, QoS, DCB and VM QoS to enforce isolation and performance guarantees

Infiniband vs. 10 GbE cs 1GbE

1Gb Ethernet: Adequate performance for many workloads
InfiniBand (32Gb and 56 Gb): Very high performance, low latency, RDMA included (SMB 3.0). Needed only when you want extreme bandwidth
10Gb Ethernet: Great performance, RDMA optional, QoS (DCB), new offloads

Hardware Offloads for Scalability & Performance

HW QoS via DCB
RDMA – For SMB storage stack only and optimized for performance
Receive Segment Coalescing (RSC)
Receive side scaling (RSS)
Virtual Machine Queue (VMQ)
Guest IPsec Task Offload (IPsecTO)
SR-IOV – For raw performance

Storage

Considerations: Cost/performance, block vs. file, Manageability, vendor preference, existing investments, approach to scaling
Storage scaling approaches: Compute and storage scale together (local SAS with storage spaces); Compute & Storage scale independently (iSCSI, FC, RDMA)

Resiliency

Infrastructure resiliency – VM is not designed to handle failures, so double up all infrastructure. Most common for the enterprise.
App-Level Resiliency – VMs designed to handle failures (e.g. guest clustering), or downtime acceptable. Lower end industry standard server, single infrastructure. Most common in cloud providers like Azure and Amazon.