For Machine Learning, It’s All About GPUs

Previously published in Forbes on December 1, 2017

Isn’t it curious that two of the top conferences on artificial intelligence are organized by NVIDIA and Intel? What do chip companies have to teach us about algorithms? The answer is that nowadays, for machine learning (ML), and particularly deep learning (DL), it’s all about GPUs.

In a previous article, I made the case to every CEO and CTO that “Machine learning allows us to make even better use of the data we have, as well as the data we don’t currently possess, and answer the questions we didn’t know we should ask.”

As more companies build AI-driven products, technology providers are responding to this demand by providing products that are computationally more powerful and easier to use and manage in production.

GPUs are driving the next wave of breakthroughs.

Why GPUs Are So Important To Machine Learning

GPUs have almost 200 times more processors per chip than a CPU. For example, an Intel Xeon Platinum 8180 Processor has 28 Cores, while an NVIDIA Tesla K80 has 4,992 CUDA cores. While a CPU core is more powerful than a GPU core, the vast majority of this power goes unused by ML applications. A CPU core is designed to support an extremely broad variety of tasks (e.g., render a webpage, drive word processors and enterprise software, manage peripherals) in addition to performing computations, whereas a GPU core is optimized exclusively for data computations. Because of this singular focus, a GPU core is simpler and has a smaller die area than a CPU, allowing many more GPU cores to be crammed onto a single chip. Consequently, ML applications, which perform large numbers of computations on a vast amount of data, can see huge (i.e., 5 to 10 times) performance improvements when running on a GPU versus a CPU.

Having recognized this fundamental fact a few years ago, the tech industry, particularly the ML crowd, has focused its efforts on taking advantage of the GPU. However, this is not a simple task. All layers of the compute stack have to be redesigned to take advantage of the GPU’s power.

Recent Developments For GPUs

NVIDIA has so far been the main provider of GPU chips for ML acceleration. The company has powered the AWS compute-optimized instances for the past year.

Furthermore, chip manufacturers are about to release chips that are architected specifically for ML from the ground up (rather than continuing to optimize GPUs, which were originally designed for graphics processing). NVIDIA is shipping the Tesla V100, which incorporates Tensor Cores designed specifically for DL, in addition to GPU cores. Google announced its Tensor Processing Unit (TPU) last year that powers its main services: Google Search, Street View, Photos and Google Translate. Finally, Intel announced this month its Nervana Neural Processor, which was also architected, in collaboration with Facebook, to optimize neural network computing.

Building The GPU Compute Stack

Having super-fast GPUs is a great starting point. In order to take full advantage of their power, the compute stack has to be re-engineered from top to bottom.

• Servers

A new category of servers needs to be built to feed the beast. This is necessary to send (and store) data to the GPU at the rate at which it is capable of consuming it, requiring up to 10x improvement in bandwidth.

NVIDIA just started shipping its DGX-1 server. Data throughput and storage have been optimized in order to take full advantage of the processing power of the eight Tesla-V100 processors included in the box.

Facebook recently announced its second generation of AI-hardware (“Big Basin”) to power its own core services: speech and text translations, photo classifiers and real-time video classification.

• Data Center

An article I wrote last month highlighted the impact of ML for cloud providers. Since then, new GPU-related developments have emerged.

Google just made its TPUs available on its compute platform.

Intel just announced its Nervana DevCloud, which is limited for the time being to research and experimentation.

Finally, a super-computing veteran of 45 years is entering the fray. Leveraging its decades of experience in high-performance computing (HPC), Cray will soon be offering its supercomputers for rent on Microsoft Azure. These servers can host a large number NVIDIA Tesla GPUs.

• Frameworks, Models And Algorithms

Optimized hardware requires optimized software. All cloud providers have optimized the major frameworks (Tensorflow, PyTorch, Caffe, MXNet) to their platform. Furthermore, GPU vendors are rewriting the major models and algorithms (NVIDIA DigitsIntel Nervana Graph) to take full advantage of the GPU’s power.

Through the GPU Open Analytics Initiative, companies such as MapD (DB, visualization) and H20 (ML) are rewriting fundamental technologies like databases and programming languages in order to eliminate data copies, which, if ignored, may significantly increase overall execution time.

Finally, some technologies have reached a degree of fidelity high enough to be offered as services: AWSGoogle and Microsoft each offer various flavors of speech recognition, translation and synthesis. Similarly, China’s Megvii’s face recognition service has become very popular.

• The Edge

For some applications, the ML models that have been trained in the data center must be computed at the edge (i.e., close to the end user). In the case of autonomous driving, for example, the car’s brain is trained in the data center but must be run in the car.

Now that machine learning has become mainstream in the data center, dedicated products are being released for edge computing. For example, NVIDIA provides the Drive PXfamily of accelerator cards that host 1-4 GPUs, as well as multiple video and other sensor inputs. They can thus power anything from simple highway driving today to fully autonomous driving in the future.

A New GPU-Driven ML Landscape

From this whirlwind survey of innovation driven by GPUs, one can anticipate increases in processing power of two to five times over the next months, from which a second wave of machine learning breakthroughs is bound to emerge, allowing us to solve a brand-new class of challenges.

 

 

Advertisements

How Machine Learning Will Disrupt The Established Cloud Providers

Previously published in Forbes on October 24, 2017

In the past few years, new categories of products have emerged thanks to the extraordinary advances in machine learning (ML) and deep learning (DL). These new techniques power product recommendations, computer-aided diagnosis in medical imaging and self-driving cars, just to name a few.

Most ML and DL algorithms require compute profiles (hardware, software, storage, networking) that are significantly different from those optimized for traditional applications. Consequently, as more and more companies develop their own ML/DL solutions and deploy them to production, the demand for the ML-optimized compute resources will grow dramatically and create opportunities for new entrants to offer solutions that compete with today’s dominant cloud providers: Amazon AWS, Microsoft Azure and Google Cloud.

The ML/DL Cloud Is Different

In an article on Mesosphere’s blog page, Edward Hsu presented the case that web applications are now primarily data-driven. Consequently, a new set of frameworks (a.k.a. stacks), namely SMACK (Spark, Mesos, Akka, Cassandra, Kafka), must replace the traditional LAMP (Linux, Apache, MySQL, PHP) stack used to build web-based applications. In my view, rather than replacing LAMP, SMACK will coexist side by side with, and feed data to, traditional web-based based frameworks, which are still needed to present nice-looking webpages and interface with mobile phones.

Yet the main point is well-taken. We need to update Marc Andreesen’s famous line about how “Software is eating the world” to “Data is eating the world.” Let’s unpack this statement and derive the consequences.

Hardware

The disruption created by machine learning and deep learning extends well beyond the software stack into chips, servers and cloud providers. This disruption is rooted in the simple fact that GPUs are much more efficient processors for ML and DL than traditional CPUs.

Up until recently, the solution was to augment traditional servers with GPU add-on cards. We are now at a point where demand for ML/DL computing is such that special-purpose servers, optimized for ML/DL compute loads, are being built.

Data centers are also being re-architected to support the extremely large amount of data consumed by ML and DL. Imagine you are designing the brains for self-driving cars. You need to process thousands and thousands of hours of video (and other such signals as GPS, gyroscopes, LIDAR) to train your algorithms. The amount of data that a Tesla on the road records in one second is a million times larger than a tweet or a post on Facebook.

ML/DL data centers thus require both huge amounts of storage and extremely high bandwidth.

Software

The software side is even more complex. A new infrastructure stack, typically using machine learning-specific frameworks such as Tensorflow (originally developed by Google) or PyTorch (originally developed at Facebook), is required to shepherd data around and manage the execution of the compute jobs. Furthermore, open-source code libraries (pandasscikit-learnmatplotlib) are used to implement the models (e.g., neural networks, data displays). These model libraries are critical because they are optimized to be both easy to use for algorithm research and offer high performance for use in production.

Finally, each vendor offers complete building blocks for specific use cases. For example, Amazon LexGoogle Cloud Speech and Microsoft Bing Speech provide speech recognition and can even recognize intent. Each has its own API and unique behavior, making the migration from one vendor to the other time-consuming.

New Entrants

In addition to the Big Three cloud providers (Amazon AWS, Microsoft Azure and Google Cloud) that have offered GPU-accelerated instances for a few years, new ML-optimized offerings have emerged:

• NVIDIA, which is already the dominant provider of GPUs that power the graphics cards that drive computer displays, recently introduced a portfolio of “purpose-built AI supercomputers” servers known as its DGX systems.

• Servers.com offers its Prisma Cloud with dedicated GPU-optimized servers.

• Rescale, one of the niche cloud providers that focuses on high-performance computing (HPC), just announced the availability of the latest generation of GPU-powered servers, along with high-bandwidth interconnect, to create high-performance multi-node clusters.

What’s At Stake

The Big Three cloud providers are the ones most immediately at risk to be disrupted by new entrants such as NVIDIA, Servers.com and Rescale. ML/DL innovation is still running at a torrid pace thanks to innovation in algorithms as well as compute efficiency. This is creating a small arms race where end users are constantly looking for the provider that can give that extra edge.

On one hand, end users are benefiting hugely from this arms race to provide the best software and hardware compute environment. On the other, this requires constant vigilance to keep abreast of the latest offerings. Even more importantly, when deploying ML/DL products to production, CEOs and CTOs need to pick the winner — or at least a future survivor — that will keep their edge for the next two to five years. This is not an easy task.

We will delve deeper into these two topics in future posts — stay tuned.

DevOps-Driven Development

It is now time to add the concept of “DevOps-Driven Development” to our repertoire.

“Test-driven” development, which originated around the same time as Extreme Programming and Agile Development, encourages us to think about testing as we architect our software and plan our tasks. Similarly, a “DevOps-Driven Development” approach, ensures that we consider operational implementation as well as deployment process during the design phase. To be clear, DevOps thinking needs to augment (and not replace) testing strategy.

Definition and Motivation

First a definition: I am using the word DevOps here as a shortcut to include both DevOps (build and deployment tools) and Ops (IT/data center Operations).

How many times have you heard “ … but it works on my machine!!” from a developer whose code was found to have a bug in the QA environment or, worse, in production? We all agree that these situations are a horrible waste of time for all involved, most of all customers. This post  thus advocates that DevOps-thinking, just as quality-thinking, must occur at the design phase and continue throughout the development of the software until the software is released to production, and even after it has been released in production.

Practicing DevOps-Driven Development

I have always advocated: “If you don’t know how to test it, you don’t know how to design it.” (Who Owns Quality? Part 3), to articulate the fact that “quality cannot be debugged out, it has to be designed in”. Similarly, if we want to know – before our customers call us – when our code crashes in Production, or becomes unusably slow, then we must build into our code the proper instrumentation and administration capabilities.

We now must add this mantra “If you don’t know how to deploy it and manage it in Production, you don’t know how to design it”.

Just like we don’t allow code to be merged into Trunk (main branch) without complete unit tests, code cannot be merged into Trunk without correct deployment scripts, release notes, and production instrumentation.

Here is a “thinking DevOps” check list:

Deployable

First of all, we must ensure that the code deploys successfully not only in Production but in all environments: Dev, QA, Stage, etc

This implies:

  • Developers write/update release notes: e.g. highlighting any changes required in the configuration of the environments: open new port, add a column in database, a new property in config files, etc
  • Developers in collaboration with DevOps team update deployment scripts, e.g. to account for a new executable, or schema changes in the database

The management of Config/Property files is beyond the scope of this blog, but I strongly recommend the “Infrastructure as code” approach: i.e. fully automating  server/image configuration for deployment and, managing configuration, deployment scripts and application property files under source code control.

Monitor-able

If we want to detect problems before our (irate) customers call us, our code needs to be monitor-able – not only at the physical server level, but also each virtual machine, service and process, as well as networking and storage systems.

Monitor-ability needs surpass keeping track of CPU load, disk space and network bandwidth. We, developers, (should) know what parameter(s) indicate when our system is mis-behaving, whether it is a queue exceeding a given size, or certain operations timing out. As a consequence, we must publish these parameters to interfaces compatible with Ops monitoring tools, of which there are several categories:

Furthermore, by making performance metrics easily observable, we ensure that each new release maintains (or improves) the performance of the prior release.

Diagnosable

Despite our best intentions, we must humbly assume that at some point our code will crash, or seriously mis-behave, and thus require troubleshooting. In the worst case, Development will be called in (usually in the wee hours of the night) to assist the Ops team. As any one who has had to figure out why a given system intermittently crashes will attest, having log files capture meaningful information prior to the incident is invaluable. Having to add logging statements after-the-fact is a painful process. Consequently, a solid Logging Hygiene is critical (and worthy of a dedicated post):

  • Log statements must be written in a format compatible with the log management system (Splunk, GrayLog2, …)
  • All log statements used during the coding and QA phase must be removed
  • Comprehensive Operations-focused logging must be added to document all operations that may fail due to environmental and data-related problems: out-of-memory, disk full, time out, user not found, access denied, etc. These are not bugs, but failures due to either environment (e.g. a server or connection is down) or incorrect data (e.g. the user has been deleted).
  • The hierarchy of logging levels must be enforced so that in normal operations log files are kept small, and conversely  meaningful information is output when troubleshooting is required
  • Log statements must include all the information necessary to bind all operations across various services that are related to a single user-level transaction (e.g. clicking on a link to a new page, adding an item to cart) – more details below in “Tunable”.

Security

This again is worthy of its own post, but code that is deployed to Production must both support the security practices implemented by the Ops team (e.g. Authentication protocols, networking infrastructure), and ensure that the code itself is secure (e.g. no SQL injection, buffer overflow, etc).

Business Continuity

Business continuity is often overlooked, but we must ensure that any persistent data is stored in a storage system that is backed up by the Ops team. In other words, if we add a new database, we’d better ask the Ops team to add it to their backup scripts.

Similarly, if our infrastructure is deployed (or even just deployable) across multiple data-centers, our code must support this though configuration.

The above requirements represent the basic DevOps requirements that any developer must address before even thinking that his/her code is ready to release. The following details additional practices that are highly recommended, but not strictly necessary.

Scalable

The code must be designed so that the Ops team can scale it in the datacenter without needing help from Development.

This may involve deploying the code to a bigger server. This implies that the code can be configured (and documented for the Ops team) to make use of the expanded resources, whether it is number of cores, RAM, threads, I/O, etc

This may also involve adding instances to a cluster. Consequently, the code must be discoverable (the load balancer must find out that a new instance has been added/subtracted), as well as cluster-aware (e.g. stateless).

Tunable

Because it is so hard to simulate all real-life user activities and behaviors in non-production environments, we must provide tools to the Ops team to tune the performance of our code through configuration rather than code deployment (e.g. size of JVM, number of threads, queue sizes, hash table size, etc).

We must thus provide the metrics to observe performance. Let’s take the example of response time: depending on the complexity of the application a user request may be handled by tens, or even hundreds of services. In order to allow the Ops team to build a timeline of the interactions between all the services involved, each log entry must carry at least one tag that identifies the root transaction that generated the request. Otherwise it is impossible to determine whether the performance degradation comes from a given service, or a unique server, or even from the network infrastructure.

The same tagging will be used to troubleshoot failures (e.g. to discover why a given service fails intermittently).

QA-able

As I mentioned in an earlier blog, QA does not stop in QA: we have to anticipate “unknown unknowns”, i.e. usage (or performance) scenarios that we have not modeled in our QA environments. By definition, there is not much we can do other than ensuring that our code is easy to trouble-shoot (see above) and that logs and associated data can be made available easily and rapidly to developers and QA team (e.g. by giving them access to the log management console).

Sometimes this requirement is more complex than it sounds, e.g. when user data must be deleted or obfuscated for privacy or security reasons. Again, this should be thought through before code is deployed.

Analytics – Growth Hacking – Usability

This last requirement stems from Marketing and Sales rather than Operations, but it is equally important since it drives revenue growth.

In most companies, marketing and sales rely on usage reports to drive new marketing campaigns, pricing, product offerings and even new features. As a consequence, any new feature must integrate with the Analytics infrastructure whether via integration with usage tracking applications (e.g. Mixpanel, Flurry, …) or simply log management consoles (Splunk, GrayLog2, …). However, I highly recommend using separate logging infrastructure for operations monitoring and for usage analytics, if only because usage analytics requires additional data that is not useful for Operations monitoring (e.g. the time a user spends on a page is extremely valuable for usage analytics but irrelevant for Operations)

Even More So for Microservices

As we migrate towards a microservices architecture, early “DevOps thinking” becomes even more critical. As the “Microservices: Four Essential Checklists when Getting Started” advises: “Microservices introduces a lot of moving parts that were previously non-existent in a monolithic system”.

What was a monolithic application running in a single virtual machine can morph into 5, 10 or even 20 microservices. Consequently, Development, DevOps and Ops must collaborate on microservices infrastructure tools: service registration, scaling up/down each service independently, health monitoring, error detection, etc. to provide visibility on the status of these 20 microservices as a whole. This challenge has even prompted dedicated product categories (SignalFx,  Nirmata, etc)

Summary

Only with a holistic approach to product architecture can we ensure customer satisfaction with software that works the first time, and all the time. Deployment and operations management concerns, just like testability, must be addressed at design time, so that these capabilities are meshed natively into the code rather than “bolted on” after the fact. Failing to do so will likely impact the delivery schedule, or worse, create outages in production.

More importantly, there is so much we can learn from observing how our code behaves in Production: operational efficiency, stability, performance, usability, that we would do a disservice to ourselves if we did not avail ourselves of this valuable information to drive further improvements to our product.

Scalable Software Architecture for a Startup

Say we are the founders of a startup and we just got a big fat check for our A-round funding. The VCs love our idea, and we all know that our app will attract millions of users in no time. This means that from day one we architect for millions of page-views per day…

But wait … do we really need to deploy Hadoop now? Do we need to design for geographical redundancy now? OR should we just build something that’s going to take us through the next 3 months, so that we can focus our energy on customer development and fine-tuning our product features? …

This is a dilemma that most startups face.

Architecting for Scale

The main argument for architecting for scale from the get-go is akin to: “do it right the first time”: we know that lots of users will be using our app, so we want to be ready when they come, and we certainly don’t want the site going down just as our product catches fire.

In addition, for those of us who have been through the pain of a complete rewrite, a rewrite is something we want to avoid at all costs: it is a complex task that is fun under the right circumstances, but very painful under time pressure, e.g. when the current version of the product is breaking under load, and we risk turning away customers, potentially for ever.

On a more modest level, working on big complex problems keeps the engineering team motivated, and working on bleeding or leading edge technology makes it easier to attract talent.

Keeping It Simple

On the other hand, keeping the technology as simple as possible allows the engineering team to be responsive to the product team during the customer development phase. If you believe, as I do, one of Steve Blank’s principles of customer development: “No Business Plan Survives First Contact with Customers”, then you need to prepare for its corollary namely: “no initial product roadmap survives first contact with customers”. Said differently, attempting to optimize the product for scale until the company has reached clear validation of its business assumptions, and product roadmap, is premature.

On the contrary, the most important qualities that are needed from the Engineering team in the early stages of the company are velocity and adaptability. Velocity, in order to reduce time-to-market, and adaptability, so that the team can rapidly adapt to feedback from “outside the building”.

Spending time designing and implementing a scalable architecture is time that is Not spent responding to customer needs. Similarly, having built a complex system makes it more difficult to adapt to changes.

Worst of all, the investment in early optimization may be all for naught: as the product evolves with customer feedback, so do the scalability constraints.

Case Study: Cloudtalk

I lived through such an example at Cloudtalk. Cloudtalk is designed as a social communication platform with emphasis on voice. The first 2 products “Cloudtalk” and “Let’s Talk” are mobile apps that implement various flavors of group messaging with voice (as well as text and other media). Predicint rapid success, Cloudtalk was designed around the highly scalable noSQL database Cassandra.

I came on board to launch “Just Sayin”, another mobile app that runs on the same backend (very astute design). Just Sayin is targeted to celebrities and allows them to cross-post voice messages to Twitter and Facebook. One of my initial tasks coming on board was to scale the app, and it was suggested that we needed it to move it to Amazon Web Services so that we can scale rapidly as more celebrities (such as Ricky Gervais) adopt our product. However, a quick analysis revealed that unlike the first two products (Let’s Talk and Cloudtalk), Just Sayin’ impact on the database was relatively light, because communications were 1-to-many (e.g. Lady Gaga to her 10M fans). Rather, in order to scale, we first needed a Content Delivery Network (CDN) so that we could feed the millions of fans the messages from their celebrities with low response time.

Furthermore, while Cassandra is a great product, it was somewhat immature at the time (stability, management tools) and consequently slowed down our development. It also took us a long time to train new engineers.

While Cassandra will have been a good choice in the long run, we would have been better served in the formative stages of the company to use more established technology like mySQL. Our velocity in developing new features, and our ability to respond to changes in product strategy would have been significantly faster.

Architecting for Scale is a Process, not an Event

A startup needs to earn the right to design for scale, by first proving that it has found a legitimate market. During this first phase adaptability and velocity are its most important attributes.

This being said, we also need to anticipate that we will need to scale the system at some point. Here is how I like to approach the problem:

  • First of all, scaling is an on-going process. Even if traffic increases dramatically over a short period of time, not all parts of the system need to be scaled at the same time. Yet, as usage increases, it is likely that any point in time, some part of the system will need to be scaled.
  • In order to avoid complete rewrites of the system, we need to break it into independent components. This allows us to redesign each component independently, and have different teams work on different problems concurrently. As a consequence, good modularization of the system is much more important early on, than designing for scale
  • Every release cycle needs to budget time and resources for redesign – including both modularization and scalability. This is just like maintenance on the Golden Gate bridge: the painters are always working; when they finish at one end, they start all over at the other end.
  • We need to treat our software architecture the same way, and budget maintenance work every release cycle: dollars, time, people. CEOs have to be trained to not only think about the “shiny features” – those that are customer-facing – but also about the “continuous improvements” of the architecture that has to be factored in every release cycle.
  • We also need to instrument the code to tell us were it is under strain. Unlike the Golden Gate bridge, we can’t always see where it’s breaking, or even rationalize it. Scaling sometimes works in mysterious ways that are not always obvious to predict.

 

In summary, designing for scale is a high-class problem, on which we only get to work once we have demonstrated true demand for our product. During this first phase, velocity and adaptability are critical, and are better served with well-understood technologies, and a well modularized design. Once our product reaches an adoption phase, then designing for scale is a continuous process that hopefully can be focused on individual modules in turn – guided by proper instrumentation of the code