Summary at the top = incredibly dense session on vSphere 5 performance….fantastic session.

Getting right into it…I love that in a session….

• Great performance maximums in vSphere 5 – 32 vCPUs, 1,000,000+ IOPs per host, etc.
  • 92-97% of native performance up to 32 vCPUs
  • CPU Scheduler Improvements — increases as much as 30%
  • vNUMA – allows NUMA-aware guest OSes to use underlying NUMA hardware more efficiently.
• vCenter — 75% increase in management ops/minute.
• HA cluster – configure 9x faster than 4.1 for a 32 host ESX cluster, 60% more VMs failover in same time.
• Network I/O Control (NIOC)
  • Traffic management for Distributed vSwitch
  • Enhancements are User-defined network resource pools, New host based replication traffic type, QoS tagging.
  • Example w/out NIOC how vMotion takes up bandwidth and impacts NFS, VM, & FT traffic.
  • Example with NIOC showing vMotion having no impact.
  • Can add QoS tags (i.e. 802.1p) so goes end to end in network environments.
• SplitRXMode
  • greatly reduces packet loss
  • new way of doing network packet receive processing in the vmkernel
    • splitting the cost of receive packet processing to multiple contexts
  • Before w/24 VMs could see up to 40% packet loss, with this now less than 10% percent.
    • Enable on a per vmnic basis.
• Multicast improvements for throughput and efficiency.
• TCP IP Stack Improvement — higher throughput with small messages & better IOPs scaling for software iSCSI
• Netflow — supported in vSphere 5.
  • Performance – monitor application performance over time
  • Capacity Planning
  • Visibility into Virtual Infrastructure Traffic
• vMotion — 25-30% improvement over 4.1, another 50% better using 2 NICs (multi-NIC vMotion)
• Storage vMotion — Live Migration & I/O Mirrorning
  • Live Migration — not dirty block tracking anymore.
    • Zero downtime maintenance.
    • Manual and automatic storage load balancing.
    • Live storage migration
• Memory Management
  • Last year with 4.1 did wide NUMA support and memory compression.
  • This year in 5.0 offering
    • vNUMA – lets know about underlying NUMA so can use fastest memory (local to the processor the VM is running on).
      • On by default for larger VMs — (8) vCPU or more.
      • Not needed for smaller VMs as ESX scheduler already keeps smaller VMs in the same NUMA domain.
    • Host Cache – use SSD storage as a cache location for VM swapped memory pages when under memory pressure.
  • New hierarchy in VMware’s memory overcommit technology
    • Transparent Page Sharing
    • Ballooning
    • Memory Compression
    • Host Cache (using SSDs)
      • Roughly 30% improvement in performance when used with other vSphere memory technologies.
    • VMs vSwap file (as a last resort)
• Storage Improvements
  • NFS support for Storage I/O Control
  • Provides…..
    • Increased Storage performance
    • Limits performance fluctuations during periods of I/O congestion
    • Increases throughput & decreases storage latency.
  • Datastore clusters
    • Group of data stores into a group.
    • Can provision VMs into the datastore cluster.
  • Storage DRS
    • Requires datastore clusters and load balances between them based on capacity and I/O.
    • What it does….
      1. Initial placement of VMs and VMDKs based on available space and I/O capacity.
      2. Load balancing between data stores in a datastore cluster via Storage vMotion based on storage space utilization.
      3. Load balancing via Storage vMotion based on I/O metrics, i.e. latency.
    • Affinity/Anti-Affinity Rules for both VMs and VMDKs.
  • VAAI — array integration
    • New primitives = thin provisioning monitoring and dead space reclamation.
    • 3 primitives last year are the big performance gains….but these are still really big deals for performance.
    • Up to 95% time reduction! 😉
    • Also saves CPU and memory resources on ESX host (so more ESX scalability).
    • Dead Space Reclamation — VM not using space (like after a Storage vMotion). VAAI gives that back to the array via a standard SCSI command.
  • 1,000,000 IOPs from a single vSphere server against an 8 engine VMAX array with a boatload of drives.
  • 10 GigE FCoE is now on par with 8 Gig FC
  • (4) vSphere hosts with dual 10 GigE Intel adapters
  • VNX Array
  • 16 streams of FCoE traffic
  • 100% sequential, 100% read, 1 MB IO size
  • 10 Gigabytes/second.
• Tier 1 Applications
• HPC Performance
  • Need high CPU counts, lots of memory, NUMA aware, etc.