IBM Edge 2016 Notes Documentation

Release master

Major Hayden

September 20, 2016

 Contents

1 Monday, September 19th 3

1.1 Trends and Directions – IBM Power Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 IBM POWER9 Technology Advanced Deep Dive . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 OpenPower Revolution in the Datacenter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2 Tuesday, September 20th 9

2.1 Getting Started with Linux Performance on IBM POWER8 . . . . . . . . . . . . . . . . . . . . . . 9
2.2 Bringing the Deep Learning Revolution into the Enterprise . . . . . . . . . . . . . . . . . . . . . . . 10

i
ii
 IBM Edge 2016 Notes Documentation, Release master

These are notes from the IBM Edge 2016 conference.

There are probably errors and omissions here. I can only type so fast, and there are certain things that are more
interesting to me than others.
I will only make notes here on the breakouts since most of the big keynote sessions are already recorded.
License: Creative Commons CC-BY-SA license

Contents 1
IBM Edge 2016 Notes Documentation, Release master

2 Contents
 CHAPTER 1

Monday, September 19th

1.1 Trends and Directions – IBM Power Systems

Speaker: Doug Balog

• Development focused on real-time analytics of large streams of data
• 78% of cloud initiatives are coordinated or fully integrated
– They have a plan for cloud in place or are using it already
– It’s all about business outcomes anyway
• Need to make it easier for customers to consume the offering
– Servers are getting more complex, but we need to do more with less
– Cloud is now a business conversation and less of a technology conversation
• New cloud models announced today for high-end use
– C-model range makes it easier to use POWER in clouds
– Reduces operating costs up to 50%
– Designed specifically for IBM AIX clients who want an on-premises cloud
• OpenStack is used to bridge the gap between private and public clouds
Guests on stage from Bosch and Staples
• Staples hybrid strategy
– Faster deployments and easier scaling on cloud
– Clouds need to connect to the backend systems of record
– Built out their own IaaS so developers could deploy their applications and iterate faster
– Developers can deploy code with a button push
• Bosch story
– Need high availability/stability
– Using in-memory databases
– Why POWER for SAP HANA?

* Better reliability/stability for enterprise IT

3
IBM Edge 2016 Notes Documentation, Release master

* Need a flexible platform to allow business to react faster

* TCO was a serious concern – POWER had a lower TCO than alternatives
– Journey to HANA – advice for others?

* Think about bigger databases

* Think about how you can react faster for customers
* Forget about spending time reconfiguring hardware
• 300 clients in the transition to HANA on POWER now
• OpenPOWER
– Five founding members started with an idea and had a shared vision
– OpenPOWER foundation has 250+ members today
– First system

* IBM has released new OpenPOWER based servers with help from NVIDIA
* First system: POWER 8 processor has 5x faster GPU throughput than Intel with new nv-link func-
tionality
– Second system

* 96TB storage available

– Third system (S812LC 1U)

* High density, high capacity compute

* Web servers, data analytics, etc
• RapidBuild Program
– Rapidly deploy mission critical applications
– Using OpenStack for management and provisioning
– Get database as a service in an integrated solution
• POWER in the public cloud
– Softlayer, Tencent, Rackspace, and Google
– FREEPOWER8 (free code to use POWER at SoftLayer today)
– Rackspace

* Member of the OpenPOWER foundation, focused on OpenCompute

* Worked with the OpenCompute community to make a POWER8-based server
* Barreleye available to purchase today!
– Google

* Bringing POWER9 to OpenCompute

• Days of Moore’s law has come to an end
– Can’t just shrink transistors over and over again
– Can’t keeping making them more dense on the chip
– Performance must come from new areas

4 Chapter 1. Monday, September 19th

 IBM Edge 2016 Notes Documentation, Release master

– Requires bringing multiple companies together to innovate on performance

– 2,500 Linux ISV’s developing on POWER
– 100K open source packages
• New announcements
– Red Hat - lots of kernel development and improvements to take advantage of new hardware capabilities
– Canonical - Released on LinuxONE, Z, Intel, POWER at the same time
– NGINX - optimizing web servers for POWER
– Mirantis
Greg from Mirantis on stage
• Mirantis/IBM working on an architecture to bring POWER 8 compute nodes into private clouds
Jeff Stuecheli from IBM on stage to talk architecture
• Yesterday’s datacenter designs
– High performance POWER mission critical stuff
– Sprawl of under-utilized servers
– Made clouds of commodity servers
– Public clouds made of mostly commodity infrastructure (driven by faster/cheaper servers)
– More reliance on software to drive performance, not hardware

* Leads to Moore’s Law problems

* Software scaling issues
• Disruptive forces
– POWER8 is 14nm, but 7nm is possible
– Intel says 4nm is exotic technology
• We used to have faster processors consuming less power each time
• Costs more and more money to get smaller and smaller (billions of dollars to build a new fab)
• Cross-correlation of data drives the need for faster processors
• Programmatic vs Cognitive
– Programmatic – you tell the computer what to do
– Cognitive – you teach the computer rather than programming it
• Workload optimized systems help us survive the end of Moore’s Law

1.2 IBM POWER9 Technology Advanced Deep Dive

Speaker: Jeff Stuecheli

• POWER9 will have multiple chips, like POWER8
– Each chip is optimized for certain workloads, like Cognitive, HPC, Cloud and Enterprise
• Core counts

1.2. IBM POWER9 Technology Advanced Deep Dive 5

IBM Edge 2016 Notes Documentation, Release master

– Memory signaling surrounds the cores

– Core size is customizable
• NVLink 2.0, CAPI 2.0 and New CAPI (25Gbit/sec)
• PCIe Gen4 with 48 lanes
• Quality of Service controls are built in for cloud/virtualized workloads
• 25Gbit/sec common link interface (will be called “Blue Link”)
– For accelerators or remote SMP
• 8 billion transistors
• Two chips:
– SMT4 Core

* 24 SMT4 cores per chip

* Linux optimized
* Commodity packaging form factor
* 8 DDR4 ports max
– SMT8 Core

* 12 SMT8 cores per chip

* PowerVM optimized
* Buffered memory attached
• Latency is greatly reduced between cores
• Efficient cores deliver 2x compute resources per socket
• Hardware-assisted garbage collection
• New interrupt architecture
– Automated partition routing for extreme virtualization
– Could be helpful for busy clouds
• Workload optimized frequency
– Manage energy between threads/cores with reduces wakeup latency
• Optical-style signaling technology delivers 25Gbit/sec
– Multi-drawer SMP connection
– NVLINK 2 GPU accelerator attach
– New CAPI accelerator attach
• New interrupt architecture
– Today: Interrupt arrives, interrupt controller can’t find the partition where the interrupt is located
– POWER9: Hardware accelerated interrupts come through with a partition ID attached to it, then broadcast
goes out to find out if the partition is executing something
– Helps a lot with densely loaded hypervisor
• Accelerators

6 Chapter 1. Monday, September 19th

 IBM Edge 2016 Notes Documentation, Release master

– Fewer hypervisor calls, user mode virtualization management

– GZIP accelerator for compression/decompression (on the CPU die, no overhead for GZIP compres-
sion/decompression)
– AES cryptography support
– True random number generation
– Data mover
• New CAPI
– Accelerator has a Power Service Layer (PSL) to reduce latency in communicating with the POWER9 chip
– Optimized CAP protocal to talk to accelerators
– PSL is moved to POWER9 die rather than accelerator – reduces latency
• 192GB/s duplex bandwidth on PCIe Gen 4 (48 lanes)
– Should see some adapters coming for Gen4 soon – especially InfiniBand adapters
• 300GB/s duplex bandwidth for Common Link (25Gbit/sec x 48 lanes)

1.3 OpenPower Revolution in the Datacenter

My laptop battery was almost dead by the time I made it to this session, but this slide deck is very similar to what
Calista Redmond presented today:
• http://www.slideshare.net/insideHPC/openpower-update

1.3. OpenPower Revolution in the Datacenter 7

IBM Edge 2016 Notes Documentation, Release master

8 Chapter 1. Monday, September 19th

 CHAPTER 2

Tuesday, September 20th

2.1 Getting Started with Linux Performance on IBM POWER8

Speaker: Steve Nasypany

Note: I stopped taking notes during this session because almost all of the useful material in this talk is already in the
slides. There are lots of links and details about which utilities you can use.

• IBM Developerworks PowerLinux Community has the latest on Linux

• lpcpu - Linux Performance Customer Profiler Utility
– Equivalent to perfpmr in AIX
– Digs up lots of diagnostic data from the server, deeper profiling
• ppc64_cpu --frequency – check to see if the CPU frequency was set very low during a PoC (previous
users may have set that)
• Use nstress on AIX/Linux to test CPU/memory/storage
• Linux pages more data out to swap than AIX does
– AIX won’t page out data to swap after boot unless memory usage is > 97%
• Use tuned to manage performance improvements
• Recent research in the Completely Fair Scheduler (CFS) scheduler shows that it needs significant optimizations
– Read The Linux Scheduler: a Decade of Wasted Cores
• Linux boots to the highest SMT setting (8) by default
• Thread counts can be controlled with ppc64_cpu command
• Linux uses virtual CPU’s (not the same as AIX’s Virtual Processors)
– AIX Virtual Processor maps to the capacity of one physical core
– Linux CPU maps to hardware thread, which could be a single core, or a single thread in a core (depending
on SMT configuration)
• Increasing SMT provides additional lanes for processing, but it won’t give better performance for a single
threaded application
• RHEL 7.1 had SMT fixes for SMT8 performance issues (exact defect fix is unknown)

9
IBM Edge 2016 Notes Documentation, Release master

• CPU accounting is different in AIX and Linux

– Linux doesn’t use calibrated PURR metrics
– If a single SMT thread is busy in Linux, you see core CPU usage as 12.5%
• CPU Metrics
– In AIX, idle time is sent back to the hypervisor
– In Linux, the steal metric is used to identify that other guests were using the CPU at the time
– High steal time means the other guests on the server are using all of the CPU time
– nmon shows steal time very well
– nmon v16 has GPU stats and CPU interrupt stats
• NUMA
– Use numastat and numactl in Linux to see how much RAM is assigned to each CPU and adjust
threads
– In POWER7, optimizing affinity could save ~ 35% performance
– In POWER8, it’s more like 10% (since POWER8 has more bandwidth)

2.2 Bringing the Deep Learning Revolution into the Enterprise

Speaker: Michael Gschwind

• Collecting data is useful but you must make sense out of it, especially in real-time
• Deep Learning is at the intersection of AI + machine learning + big data
• What is deep learning?
– Things you do without thinking about it – like recognizing a bicycle
– Bicycles come in many forms and it’s difficult to teach a computer how to recognize them
– Humans can deal with complexity/variance easily, but traditional recognition technology cannot
– Deep learning involves building a model based on patterns, and then allow you to take an action on that
(and predict outcomes)
• Applications
– Automotive/transportation: driverless cars
– Security/Safety: facial/object recognition
– Consumer web/mobile/retail: natural language processing, recommendations
– Medicine/biology: diagnostic assistance and drug discovery
– Broadcast/media: Captioning, recommendations
• Computing capacity has been a big limiter in deep learning, but accelerators are helping
• Based on artificial neural networks (ANNs)
– Inspired by biological neural networks in the brain
– A neuron is a basic cognitive unit and you need a lot of them to begin making decisions

10 Chapter 2. Tuesday, September 20th

 IBM Edge 2016 Notes Documentation, Release master

– Expressed in computers as multiple layers, nodes, weights and activation functions (what to do when a
threshold is crossed)
– Recognizing something takes multiple steps in animals and in computers
• Training a neural network
– Called stochastic gradient descent (SGD)
– Shows how well a trained network is
– Involves taking a lot of derivatives to find “valleys”
• Middleware exists to help with the mathematics and training of a neural network, such as Caffe, TensorFlow,
CNTK, and torch
• Two-tiered workload
– Training: extract a model from historical data
– Inference: classify new objects based on the model
• Computers will return results with a probability attached
• More training improves the probability of being right
• Pre-packaged for POWER systems: http://ibm.biz/power-mldl

2.2. Bringing the Deep Learning Revolution into the Enterprise 11