VMware Experts Program Day 3
This is Day 3 of the VMware Experts Program Big Data, Scientific & Engineering Workloads at VMware Corporate Headquarters, Palo Alto, Ca.
Compute Accelerators for HPC, ML, and Big Data on vSphere
ZivKalmanovich
The problem with a Paravirtual device is performance?You have another layer to go through
GPU Compute on vSphere with DirectPath IO
Ability to share a single GPU across Virtual Machines
Containers and Big Data
Ben Corrie
Did a research project many years ago to see if we could make VMs ?Look and Smell like containers
People are still figuring out how to get the benefits of containers for their existing?applications
Turning Policy into plumbing
Containers and VMs are complementary
When you put multiple containers in a VM they are in a failure domain
What is a container?
- An Executable process
- Designed to be one process
- Resource constraints/Private Namespace
- Binary dependencies: Application runtime, OS
- A shared Linux kernel for running the executable
- Ephemeral and persistent storage layer
Difference between a hypervisor and a container host, A hypervisor is running everything
Containers force you to think about state management in a way that is really helpful
A transactional container only consumes resources when it’s running
Pods: Ability to tie multiple containers into a single ??Unit of Scale?
DAWN: Infrastructure for Usable Machine Learning
Matei Zaharia, Stanford
It?s the golden age of data
Hidden Technical Debt in Machine Learning Systems from Google
Training Data is the Key to AI –
Image search, Speech, Games: Labeled Training Data is (Relatively) easy to obtain
How are we handling performance: End-to-end compilers: WELD, Delite
The main way developers are productive is by composing existing?libraries
For data-intensive apps, data movement costs dominates on modern hardware
Machine learning systems can do much more to support user applications end-to-end
Successful systems need to span whole software stack from infrastructure to data to algorithms
Reference Architectures for High-Performance Computing
Mohan Potheri
Since HPC are long-running jobs it?s important to leverage vSphere to mitigate impact of infrastructure failures
You can provide networking at a VM level with Software-defined networking.
For HPC environment you are able to combine all your resources
Secure multi-tenancy with NSX
Nearly every IT Component contributes to application performance
A lack of visibility can lead to alert storms that drain the productivity of HPC systems
Workload Management for vSphere
Jared Rosoff
Workloads are getting more complicated
A Workload is not just a VM
These workloads are not tied to a host of a cluster – adding a new global view
Tagging would be great if everything had the right tag on it.
Workloads access would be role based
Machine Learning Using Virtualized GPUs in vSphere? Uday Kurkure
Two solutions to access GPU in vSphere VMWare DirectPathIO or Nvidia Grid vCPU
Speaking on GPUs Overhead in respect to native is only 4% in both solutions
GPUs: More Silicon is devoted to increasing the numbers of ALUs
GPUs outperform CPUs on ML Workloads
GPUs is sitting on PCIe bus
One VM with one vGPU can be scaled to four VMs with one vGPU each
You really run multiple virtual machines on P40
Virtualized GPUs deliver near bare-metal performance for ML workloads in VMware vSphere
Here are some links to the first day
Michael Corey
*** All Posts are Mine and Not those of my Employer***
LicenseFortress – Real Time Oracle License Compliance Alerting and Management with a Financial Guarantee.
My Blog: http://michaelcorey.com/
My Personal Twitter Account:?Michael_Corey
Columnist for the Big Data Quarterly. <Click Here to Subscribe Big Data Quarterly>
Buy at VMWarePress!
