VMworld 2014 Europe, day two

(Note! This blog post is a personal notepad for the day, so it might be a little bit confusing to read :))

Day started with a General Session. Lot's of hype and marketing jargon, also some new product releases (most of them already done at VMworld US)

Good list of new releases can be found here.

Lots or rebranding

STO 2197 - Storage DRS: Deepdive and Best Practices:

Storage DRS basic functionality:

• Ease of Storage Management
• Initial placement (automatically select datastore)
• Out of Space Avoidance
• IO Load Balancing
• Virtual Disk Affinity (Anti-Affinity)
• Datastore maintenance mode

Storage DRS can handle space over-commit either in Hypervisor level or in Datastore level through VASA integraion..
Best practice: never combine those!

Many new features are in vSpere 6.0

• Brake large I/O to 32KB
• Reservations to storage I/O
• If reservation cannot be satisfied, Storage DRS will kick in.
• Fine grain automatic actions

Best practices:

• avoid mixing of vSphere LUNs and non-vSphere LUNs on the same physical storage
• Configure host IO queue size with highest allowed value
• Keep congestion threshold conservatively high
• More datastores = better IO control, bigger datastore = better storage management

After that session, I went to do VCP5-DCV test, and...... Finally, achieved VCP status! :)

NET1743 - VMware NSX - A Technical Deep Dive

VMware is talking a lot about NSX, but at least in Finland, when can you buy it? How much does it cost? Technology is very interesting, might even say very cool.

I'm not going to cover this NSX  session, since lot of that information is already available on VMware blogs and documents. One link: www.vmware.com/products/nsx/resources.html

NET1586 - Advanced Network Services with NSX

NSX: offers:

• Switching / DCHP server-or-relay / DNS
• Routing / NAT
• Firewalling
• Load Balancing
• L2 / L3 VPN

Firewall is centrally managed, distributed FW. Works in front of VMNIC.

Security within the same IP-subnet.

Buzzword: micro-segmentation!

FW rules are not dependent on IP-addresses, you can use object names VMware to build rules. And those rules can be dynamic.

INF1469 - Extreme Performance Series: Monster VM Performance

Do idle vCPU cause co-stop? No they don't, there is no co-stop overhead on idle vCPUs.

Single threaded app -> multiple vCPU -> might even hurt performance

Over-provisioning vCPUs is not free -> Only configure as many vCPUs as the app needs!

Is %CSTP (co-stop) indicating a performance issue?

• %RDY + %CSTP = less than a few percent per vCPU should be ok

To avoid co-stop, address the slowest vCPU

• co-stop begins if Progress(vCPU) - Progress(vCPU_slowest) > costop_threshold
• Possible causes for vCPU_slowest
• High ready time (%RDY)
• High wait time (%VMWAIT = %WAIT - %IDLE)
• Snapshot, delta-disks, linked-clones
• VMM-level synchronization (memory manipulation: page sharing, NUMA remapping etc)

NUMA

• non-uniform memory access system architecture
• each node consists of CPU cores and memory
• pCPU can cross NUMA nodes, but at a performance cst

vNUMA

• Problem
• BIG VM has more vCPUs than # cores per NUMA node
• Guest app/OS has no knowledge of underlying NUMA
• Solutions:
• Expose virtual NUMA to VM
• Guest app/os achieves optimal placement of processes and memory

vNUMA is enabled:

• HW version higher than 8
• VM is wide (More vCPU per VM than # cores per NUMA node)
• Hyperthreading does not count
• by default, is configured with VMs with more than 8vCPU

How to ensure optimal vNUMA?

• Size VM to be multiples of pNUMA size
• 8/16/24... vCPU VMs for 8core/node
• 10/20/30... vCPU VMs for 10 cores/node

Virtual Sockets & Virtual Cores per Socket

• Add sockets, don't cores (if licensing allows you to)

Memory overcommitting = not recommended. Leave 10% room for ESXi.

• On huge VMs (large memory), consider using memory reservation to avoid creating huge swap files (.vswp) on datastore

Now the day is over (1800 local time), and it's time to celebrate my VCP status!