Tech.Ed SVR315 Windows 2008 Scalability and Performance

Posted: November 20th, 2008 under Uncategorized.

My notes from Tech.Ed session

Windows 2008 is the final x86 Microsoft OS. R2 is 64bit only.

File Server SMB 2.0
SMB2.0 (Remote File Sharing protocol v2)
Introduced in Vista and W2K8
SMB 1 is serialised
SMB2 can package up right requests and receive multiple write responses
Also support many more file locks than in SMB2.0

Next Gen TCP/IP stack
Rewritten from scratch
IPv4 and IPv6 native on same stack
W2K3 has emulated IPv6
Intelligent automated tuning of TCP receive window size (MTU)
Advanced congestion control for better throughput
Allows network stack to leverage multiple cores W2K3 had a single stack for all cores

Receive Size Scaling
In W2K3 there is no benefit in having multi CPU machines with a single NIC under heavy load
In W2K8 all processors can be used simultaneously for incoming traffic
Incoming connection distributed across available processors

Virtualisation
Hyper V hypervisor is 500K on disk
Supports Intel VT and AMD-v extensions
Small attack surface and no driver model
- drivers and pushed to the root VM (above Hypervisor) that the child VMs use.
Virtualises guest memory and schedules virtual processors
Support multiprocessor and x64 clients
Maximum 32 cores
OS is "enlightenments" in WS08 guests to enhance operations in VM environments. OS knows it is been virtualised and changes internal architecture to be optimum for virtualised

NUMA
Non uniform Memory Access
In a non NUMA architecture the writes from memory are tired to specific processors due to bus architecture. so in a 4 core machine if P3 needs to initiate a disk right P1 needs to pause what its doing to facilitate the write. NUMA resolves this.

I/O processing can be made to happen on the same request
4 major host bus adapter vendors have committed to putting NUMA I/O into the products
At least 10% reduction in CPU utilisation for I/O intensive workloads

Remote Large File Copy on WAN (benchmark with 100ms latency)

Huge benefits e.g. 6Mb file takes 4 seconds on W2K8 SMB2 takes 25 seconds on XP/W2K3 SMB1 system
The larger the file the better the gain
W2K3 networking can support more concurrent connections and more throughput.
Typically increase of 10%
More cores / CPUs increases performance

Some Sizing Metrics from Virtualised Microsoft.com 

Microsoft.com handles 10% more requests when migrated to Hyper-V W2K8 + IIS7
100M extra requests processed a day or 3 billion extra requests per month on same physical hardware

SAP 5% increase in performance

Terminal Server Workload

Significant increase in high end scalability
40%+ increase in users supported per system when going from 16 to 32 cores
5% increase at 16 cores
This relates to receive size scaling – RSS due to multiprocessor / network impotents

www.microsoft.comWhdc/system/sysperf/perf_tun_srv.mspx

NUMA I/O performance
16 Core 20% improvement over W2K3

W2K8 uses 10% less power than systems running at 25% utilisation out of the box
Bigger increases in efficiency can be achieved by tuning

W2K8 R2 expect even more power efficiency – feature called core parking
Machine will detect load and power of cores

Windows Firewall
On by Default in 2K8 5% overhead

No Comments »

No comments yet.

RSS feed for comments on this post. TrackBack URL

Leave a comment