No Huddle Offense

"Individual commitment to a group effort-that is what makes a team work, a company work, a society work, a civilization work."

Example 1: Intelligent Orchestration & Scheduling with Kubernetes

September 18th, 2016 • Comments Off on Example 1: Intelligent Orchestration & Scheduling with Kubernetes

In the last blog I suggested that analytical capabilities need to move to the core of resource managers. This is very much needed for autonomous controlled large scale systems which figure out the biggest chunk of decisions to be made themselves. While the benefits from this might be obvious, the how to inject the insights/intelligence back into the resource manager might not be. Hence this blog post series documenting a bit how to let systems (they are just examples – trying to cover most domains :-)) like Kubernetes, OpenStackMesos, YARN and OpenLava make smarter decisions.

Background

The blog posts are going to cover some generic concepts as well as point to specific documentation bits of the individual resource managers. Some of this is already covered in past blog posts but to recap let’s look at the 5(+1) Ws  for resource managers decision making (click to skip to the technical details):

Decision most of the time cannot be made generic – e.g. decisions made in HPC/HTC systems do not necessarily apply to a telco environments in which the workloads and resource are different. Hence the context of workloads and resource in place play a huge role. Ultimately Analytics which embraces the context (in all sorts and forms: deep/machine learning, statistical modelling, artificial intelligence, …) of the environment can drive the intelligence in the decision making through insights. This can obviously in multiple places/flows (see the foreground and background flow concepts here) and ultimately enables autonomous control.

Enhancing Kubernetes

For the Kubernetes examples let’s focus on a crucial decision point – doing the initial placement of a workloads (aka a POD in Kubernetes language) in a cluster. Although much of today’s research focuses on initial placement I’d urge everybody not to forget about all the other decisions that can be made more intelligent.

Like most Orchestrators and Schedulers Kubernetes follows a simple approach of filtering and ranking. After shortlisting possible candidates, the first step involves filtering those resource which do not meet the workloads demands. The second step involves prioritization (or ranking) of the resources best suited.

This general part is described nicely in the Kubernetes documentation here: https://github.com/kubernetes/kubernetes/blob/master/docs/devel/scheduler.md

This filtering part is mostly done based on capacities, while the second can involve information like the utilization. If you want to see this code have a look at the generic scheduling implementation: here. The available algorithms for filtering (aka predicates) and prioritization can be found here. The default methods that Kubernetes filters upon can be seen here: here – the default prioritization algorithms here: here. Note that weights can be applied to the algorithms based on your own needs as a provider. This is a nice way to tune and define how the resource under the control of the provider can be used.

While the process and the defaults already do a great job – let’s assume you’ve found a way on when and how to use an accelerator. Thankfully like most scheduling systems the scheduler in Kubernetes is extendable. Documentation for this can be found here. 3 ways are possible:

  1. recompile and alter the scheduler code,
  2. implement your own scheduler completely and run it in parallel,
  3. or implement an extension which the default scheduler calls when needed.

The first option is probably hard to manage in the long term, the second option requires you to deal with the messiness or concurrency while the third option is interesting (although adds latency to the process of scheduling due to an extra HTTP(s) call made). The default scheduler can basically call an external process to either ‘filter’ or ‘prioritize’. In the first case a list of possible candidate hosts is returned, in the the second case a prioritized list if returned. Now unfortunately the documentation get’s a bit vague, but luckily some code is available from the integration tests. For example here you can see some external filtering code, and here the matching prioritization code. Now that just needs to be served up over HTTP and you are ready to go, next to adding some configurations documented here.

So now an external scheduler extension can make a decisions if an accelerator should be assigned to a workload or not. The intelligent decision implemented in this extender could e.g. decide if an SR-IOV port is needed based on a bandwidth requirement, or if it is even a good idea to assign a Accelerator to a workload par the previous example.

Corrections, feedback and additional info are more then welcome. I’ve some scheduler extender code running here – but that is not shareable yet. I will update the post once I’ve completed this. In the next posts OpenStack (e.g. service like Nova, Watcher, Heat and Neutron), Mesos (how e.g. allocator modules can be used to inject smarts) and OpenLava (for which e.g. elims can be used to make better scheduling decisions) and obviously others will be introduced 🙂

 

Insight driven resource management & scheduling

July 25th, 2016 • Comments Off on Insight driven resource management & scheduling

Future data center resource and workload managers – and their [distributed]schedulers – will require a new key integrate capability: analytics. Reason for this is the the pure scale and the complexity of the disaggregation of resources and workloads which requires getting deeper insights to make better actuation decisions.

For data center management two major factors play a role: the workload (processes, tasks, containers, VMs, …) and the resources (CPUs, MEM, disks, power supplies, fans, …) under control. These form the service and resource landscape and are specific to the context of the individual data center. Different providers use different (heterogeneous) hardware (revisions) resource and have different customer groups running different workloads. The landscape overall describes how the entities in the data center are spatially connected. Telemetry systems allow for observing how they behave over time.

The following diagram can be seen as a metaphor on how the two interact: the workload create a impact on the landscape. The box represent a simple workload having an impact on the resource landscape. The landscape would be composed of all kind of different entities in the data center: from the air conditioning facility all the way to the CPU. Obviously the model taken here is very simple and in real-life a service would span multiple service components (such as load-balancers, DBs, frontends, backends, …). Different kinds of workloads impact the resource landscape in different ways.

landscape_gravity

(Click to enlarge)

Current data center management systems are too focused on understanding resources behavior only and while external analytics capabilities exists, it becomes crucial that these capabilities need to move to the core of it and allow for observing and deriving insights for both the workload and resource behavior:

Deriving insights on how workloads behave during the life-cycle, and how resources react to that impact, as well as how they can enhance the service delivery is ultimately key to finding the best match between service components over space and time. Better matching (aka actually playing Tetris – and smartly placing the title on the playing field) allows for optimized TCO given a certain business objective. Hence it is key that the analytical capabilities for getting insights on workload and resource behavior move to the very core of the workload and resource management systems in future to make better insightful decisions. This btw is key on all levels of the system hierarchy: on single resource, hosts, resource group and cluster level.

Note: parts of this were discussed during the 9th workshop on cloud control.

Graph stitching

January 2nd, 2016 • Comments Off on Graph stitching

Graph stitching describes a way to merge two graphs by adding relationships/edges between them. To determine which edges to add, a notion of node types is used (based on node naming would be easy :-)). Nodes with a certain type can be “stitched” to a node with a certain other type. As multiple mappings are possible, multiple result/candidate graphs are possible. A good stitch is defined by:

So based on node types two graphs are stitched together, and than a set of candidate result graphs will be validated, to especially satisfy the second bullet.

Let’s use an example to explain this concept a bit further. Assume the electrical “grid” in a house can be described by a graph with nodes like the power outlets and fuses, as well as edges describe the wiring. Some home appliances might be in place and connected to this graph as well. Hence a set of nodes describing for example a microwave (the power supply & magnetron), are in this graph as well. The edge between the power supply and the power outlet describe the power cable. The edge between the power supply and the magnetron is the internal cabling within the microwave. This graph can be seen in the following diagram.

(Click on enlarge)

(Click on enlarge)

The main fuse is connected to the fuses 1 & 2. Fuse 1 has three connected power outlets, of which outlet #2 is used by the microwave. Fuse 2 has two connected power outlets. Let’s call this graph the container from now on.

Now let’s assume a new HiFi installation (consisting of a blu-ray player and an amplifier) needs to be placed within this existing container. The installation itself can again be described using a simple graph, as shown in the following diagram.

(Click to enlarge)

(Click to enlarge)

Placing this request graph into the container graph now only requires that the power supplies of the player and amplifier are connected to the power outlets in the wall using a power cord. Hence edges/relationships are added to the container to stitch it to the request. This is done using the following mapping defition (The power_supply and power_outlet are values for the attribute “type” in the request & container graph):

{
    "power_supply": "power_outlet"
}

As there is more than one possible results for stitching two graphs, candidates (there are 2 power supplies and 5 power outlets in the mix) need to be examined to see if they make sense (e.g. the fuse to which the microwave is connected might blow up if another “consumer” is added). But before getting to the validation, the number of candidates graphs should be limited using conditions.

For example the HiFi installation should be placed in the living room and not the kitchen. Hence a condition as follows (The power outlet nodes in the container graph have an attribute which is either set to ‘kitchen’ or ‘living’) can be defined:

condition = {
    'attributes': [
        ('eq', ('bluray_p', ('room', 'living'))),
        ('eq', ('amp_p', ('room', 'living'))),
    ]
}

Also the amplifier should not be placed in the kitchen while the blu-ray player is placed in the living room. Hence the four nodes describing the request should share the same value for the room attribute. Also it can be defined that the power supplies of the player & amplifier should not be connected to the same power outlet:

condition = {
    'compositions': [
        ('share', ('room', ['amp', 'amp_p', 'bluray', 'bluray_p'])),
        ('diff', ('amp_p', 'bluray_p'))
    ]
}

This already limits the number of candidate resulting graphs which need to validated further. During validation it is determined if a graph resulting out of a possible stitching falls under the definition of a good stitch (see earlier on). Within the container – shown early – the nodes are ranked – red indicating the power outlet or fuse heavy loaded; while green means the power outlet/fuse is doing fine. Now let’s assume no more “consumers” should be added to the second outlet connected to the first fuse as the load (rank) is to high. The high load might be caused by the microwave.

All possible candidate graphs (given the second condition described earlier) are shown in the following diagram. The titles of the graphs describe the outcome of the validation, indicating that adding any more consumers to outlet_2 will cause problems:

(Click to enlarge)

(Click to enlarge)

The container and request are represented as shown earlier, while the stitches for each candidate resulting graph are shown as dotted lines.

Graph stitcher is a simple tool implements a simple a stitching algorithm which generates the possible graphs (while adhering all kinds of conditions). These graphs can than be validated further based on different validators. For example by looking at number of incoming edges, node rank like described before, or any other algorithm. The tool hence can be seen as a simple framework (with basic visualisation support) to validate the concepts & usefulness of graph stitching.

American Football Game Analysis

October 30th, 2014 • Comments Off on American Football Game Analysis

I’ve been coaching American Football for a while now and it is a blast standing on the sideline during game day. The not so “funny” part of coaching however – especially as Defense Coordinator – is the endless hours spend on making up stats of the offensive strategy of the opponent. Time to save some time and let the computer do the work.

I’ve posted about how you could use suricate in a sports data setup past. The following screen shot show the first baby steps (On purpose not the latest and greatest – sry 🙂 ) of analyzing game data using suricate with python pandas and scikit-learn for some clustering. The 3D plot shows Down & Distance vs Run/Pass plays. This is just raw data coming from e.g. here.

The colors of the dots actually have a meaning in such that they represent a clustering of many past plays. The clustering is done not only on Down & Distance but also on factors like field position etc. So a cluster can be seen as a group of plays with similar characteristics for now. These clusters can later be used to identify a upcoming play which is in a similar cluster.

(Click to enlarge)

(Click to enlarge)

The output of this python script stores processed data back to the object store of suricate.

One of the new features of suricate is template-able dashboards (not shown in past screenshot). Which basically means you can create custom dashboards with fancy graphics (choose you poison: D3, matplotlib, etc):

(Click to enlarge)

(Click to enlarge)

Again some data is left out for simplicity & secrecy 🙂

Making use of the stats

One part is understanding the stats as created in the first part. Secondly acting upon it is more important. With Tablets taking on sidelines, it is time to do the same & take the stats with you on game day. I have a simple web app sitting around in which current ball position is entered and some basic stats are shown.

This little web application does two things:

  1. Send a AMQP msg with the last play information to a RabbitMQ broker. Based on this new message new stats are calculated and stored back to the game data. This works thanks to suricate’s streaming support.
  2. Trigger suricate to re-calculate the changes of Run-vs-Pass in an upcoming play.

The webapp is a simple WSGI python application – still the hard work is carried out by suricate. Nevertheless the screenshot below shows the basic concept:

(Click to enlarge)

(Click to enlarge)

Running a distributed native-cloud python app on a CoreOS cluster

September 21st, 2014 • 1 Comment

Suricate is an open source Data Science platform. As it is architected to be a native-cloud app, it is composed into multiple parts:

Up till now each part was running in a SmartOS zone in my test setup or run with Openhift Gears. But I wanted to give CoreOS a shot and slowly get into using things like Kubernetes. This tutorial hence will guide through creating: the Docker image needed, the deployment of RabbitMQ & MongoDB as well as deployment of the services of Suricate itself on top of a CoreOS cluster. We’ll use suricate as an example case here – it is also the general instructions to running distributed python apps on CoreOS.

Step 0) Get a CoreOS cluster up & running

Best done using VagrantUp and this repository.

Step 1) Creating a docker image with the python app embedded

Initially we need to create a docker image which embeds the Python application itself. Therefore we will create a image based on Ubuntu and install the necessary requirements. To get started create a new directory – within initialize a git repository. Once done we’ll embed the python code we want to run using a git submodule.

$ git init
$ git submodule add https://github.com/engjoy/suricate.git

Now we’ll create a little directory called misc and dump the python scripts in it which execute the frontend and execution node of suricate. The requirements.txt file is a pip requirements file.

 
$ ls -ltr misc/
total 12
-rw-r--r-- 1 core core 20 Sep 21 11:53 requirements.txt
-rw-r--r-- 1 core core 737 Sep 21 12:21 frontend.py
-rw-r--r-- 1 core core 764 Sep 21 12:29 execnode.py 

Now it is down to creating a Dockerfile which will install the requirements and make sure the suricate application is deployed:

 
$ cat Dockerfile
FROM ubuntu
MAINTAINER engjoy UG (haftungsbeschraenkt)

# apt-get stuff
RUN echo "deb http://archive.ubuntu.com/ubuntu/ trusty main universe" >> /etc/apt/sources.list
RUN apt-get update
RUN apt-get install -y tar build-essential
RUN apt-get install -y python python-dev python-distribute python-pip

# deploy suricate
ADD /misc /misc
ADD /suricate /suricate

RUN pip install -r /misc/requirements.txt

RUN cd suricate && python setup.py install && cd ..

Now all there is left to do is to build the image:

 
$ docker build -t docker/suricate .

Now we have a docker image we can use for both the frontend and execution nodes of suricate. When starting the docker container we will just make sure to start the right executable.

Note.: Once done publish all on DockerHub – that’ll make live easy for you in future.

Step 2) Getting RabbitMQ and MongoDB up & running as units

Before getting suricate up and running we need a RabbitMq broker and a Mongo database. These are just dependencies for our app – your app might need a different set of services. Download the docker images first:

 
$ docker pull tutum/rabbitmq
$ docker pull dockerfile/mongodb

Now we will need to define the RabbitMQ service as a CoreOS unit in a file call rabbitmq.service:

 
$ cat rabbitmq.service
[Unit]
Description=RabbitMQ server
After=docker.service
Requires=docker.service
After=etcd.service
Requires=etcd.service

[Service]
ExecStartPre=/bin/sh -c "/usr/bin/docker rm -f rabbitmq > /dev/null ; true"
ExecStart=/usr/bin/docker run -p 5672:5672 -p 15672:15672 -e RABBITMQ_PASS=secret --name rabbitmq tutum/rabbitmq
ExecStop=/usr/bin/docker stop rabbitmq
ExecStopPost=/usr/bin/docker rm -f rabbitmq

Now in CoreOS we can use fleet to start the rabbitmq service:

 
$ fleetctl start rabbitmq.service
$ fleetctl list-units
UNIT                    MACHINE                         ACTIVE  SUB
rabbitmq.service        b9239746.../172.17.8.101        active  running

The CoreOS cluster will make sure the docker container is launched and RabbitMQ is up & running. More on fleet & scheduling can be found here.

This steps needs to be repeated for the MongoDB service. But afterall it is just a change of the Exec* scripts above (Mind the port setups!). Once done MongoDB and RabbitMQ will happily run:

 
$ fleetctl list-units
UNIT                    MACHINE                         ACTIVE  SUB
mongo.service           b9239746.../172.17.8.101        active  running
rabbitmq.service        b9239746.../172.17.8.101        active  running

Step 3) Run frontend and execution nodes of suricate.

Now it is time to bring up the python application. As we have defined a docker image called engjoy/suricate in step 1 we just need to define the units for CoreOS fleet again. For the frontend we create:

 
$ cat frontend.service
[Unit]
Description=Exec node server
After=docker.service
Requires=docker.service
After=etcd.service
Requires=etcd.service

[Service]
ExecStartPre=/bin/sh -c "/usr/bin/docker rm -f suricate > /dev/null ; true"
ExecStart=/usr/bin/docker run -p 8888:8888 --name suricate -e MONGO_URI=<change uri> -e RABBITMQ_URI=<change uri> engjoy/suricate python /misc/frontend.py
ExecStop=/usr/bin/docker stop suricate
ExecStopPost=/usr/bin/docker rm -f suricate

As you can see it will use the engjoy/suricate image from above and just run the python command. The frontend is now up & running. The same steps need to be repeated for the execution node. As we run at least one execution node per tenant we’ll get multiple units for now. After bringing up multiple execution nodes and the frontend the list of units looks like:

 
$ fleetctl list-units
UNIT                    MACHINE                         ACTIVE  SUB
exec_node_user1.service b9239746.../172.17.8.101        active  running
exec_node_user2.service b9239746.../172.17.8.101        active  running
frontend.service        b9239746.../172.17.8.101        active  running
mongo.service           b9239746.../172.17.8.101        active  running
rabbitmq.service        b9239746.../172.17.8.101        active  running
[...]

Now your distributed Python app is happily running on a CoreOS cluster.

Some notes