User Guide¶

ConPaaS currently contains the following services:

Two Web hosting services respectively specialized for hosting PHP and JSP applications;
MySQL offering a multi-master replicated load-balanced database service;
Scalarix service offering a scalable in-memory key-value store;
MapReduce service providing the well-known high-performance computation framework;
TaskFarming service high-performance batch processing;
Selenium service for functional testing of web applications;
XtreemFS service offering a distributed and replicated file system;
HTC service providing a throughput-oriented scheduler for bags of tasks submitted on demand;
Generic service allowing the execution of arbitrary applications.

ConPaaS applications can be composed of any number of services. For example, a bio-informatics application may make use of a PHP and a MySQL service to host a Web-based frontend, and link this frontend to a MapReduce backend service for conducting high-performance genomic computations on demand.

Usage overview¶

Web-based interface¶

Most operations in ConPaaS can be done using the ConPaaS frontend, which gives a Web-based interface to the system. The front-end allows users to register (directly with ConPaaS or through an external Identification Provider at Contrail), create services, upload code and data to the services, and configure each service.

The Dashboard page displays the list of services currently active in the system.
Each service comes with a separate page which allows one to configure it, upload code and data, and scale it up and down.

Command line interfaces¶

All the functionalities of the frontend are also available using a command-line interface. This allows one to script commands for ConPaaS. The command-line interface also features additional advanced functionalities, which are not available using the front-end. (The use of external Identification Provider at Contrail is not yet available from the command-line interface.)

It exists two command line clients: cpsclient.py and cps-tools.

cpsclient.py

Installation and configuration: see Installing and configuring cpsclient.py.

Command arguments:

$ cpsclient.py usage

Available service types:

$ cpsclient.py available

Service command specific arguments:

$ cpsclient.py usage <service_type>

Create a service:

$ cpsclient.py create <service_type>

List services:

$ cpsclient.py list

cps-tools

Installation and configuration:: see Installing and configuring cps-tools.

Command arguments:

$ cps-tools --help

Available service types:

$ cps-tools service get_types
$ cps-service get-types

Service command specific arguments:

$ cps-tools <service_type> --help
$ cps-<service_type> --help

Create a service:

$ cps-tools service create <service_type>
$ cps-tools <service_type> create
$ cps-<service_type> create

List services:

$ cps-tools service list
$ cps-service list

List applications:

$ cps-tools application list
$ cps-application list

List clouds:

$ cps-tools cloud list
$ cps-cloud list

Controlling services using the front-end¶

The ConPaaS front-end provides a simple and intuitive interface for controlling services. We discuss here the features that are common to all services, and refer to the next sections for service-specific functionality.

Create a service.: Click on “create new service”, then select the service you want to create. This operation starts a new “Manager” virtual machine instance. The manager is in charge of taking care of the service, but it does not host applications itself. Other instances in charge of running the actual application are called “agent” instances.
Start a service.: Click on “start”, this will create a new virtual machine which can host applications, depending on the type of service.
Rename the service.: By default all new services are named “New service”. To give a meaningful name to a service, click on this name in the service-specific page and enter a new name.
Check the list of virtual instances.: A service can run using one or more virtual machine instances. The service-specific page shows the list of instances, their respective IP addresses, and the role each instance is currently having in the service. Certain services use a single role for all instances, while other services specialize different instances to take different roles. For example, the PHP Web hosting service distinguishes three roles: load balancers, web servers, and PHP servers.
Scale the service up and down.: When a service is started it uses a single “agent” instance. To add more capacity, or to later reduce capacity you can vary the number of instances used by the service. Click the numbers below the list of instances to request adding or removing servers. The system reconfigures itself without any service interruption.
Stop the service.: When you do not need to run the application any more, click “stop” to stop the service. This stops all instances except the manager which keeps on running.
Terminate the service.: Click “terminate” to terminate the service. At this point all the state of the service manager will be lost.

Controlling services using the command-line interfaces¶

Command-line interfaces allow one to control services without using the graphical interface. The command-line interfaces also offer additional functionalities for advanced usage of the services. See Installing and configuring cpsclient.py to install it.

List all options of the command-line tool.

$ cpsclient.py help

Create a service.

$ cpsclient.py create php

List available services.

$ cpsclient.py list

List service-specific options.

# in this example the id of our service is 1
$ cpsclient.py usage 1

Scale the service up and down.

$ cpsclient.py usage 1
$ cpsclient.py add_nodes 1 1 1 0
$ cpsclient.py remove_nodes 1 1 1 0

The credit system¶

In Cloud computing, resources come at a cost. ConPaaS reflects this reality in the form of a credit system. Each user is given a number of credits that she can use as she wishes. One credit corresponds to one hour of execution of one virtual machine. The number of available credits is always mentioned in the top-right corner of the front-end. Once credits are exhausted, your running instances will be stopped and you will not be able to use the system until the administrator decides to give additional credit.

Note that every service consumes credit, even if it is in “stopped” state. The reason is that stopped services still have one “manager” instance running. To stop using credits you must completely terminate your services.

Tutorial: hosting WordPress in ConPaaS¶

This short tutorial illustrates the way to use ConPaaS to install and host WordPress (http://www.wordpress.org), a well-known third-party Web application. WordPress is implemented in PHP using a MySQL database so we will need a PHP and a MySQL service in ConPaaS.

Open the ConPaaS front-end in your Web browser and log in. If necessary, create yourself a user account and make sure that you have at least 5 credits. Your credits are always shown in the top-right corner of the front-end. One credit corresponds to one hour of execution of one virtual machine instance.
Create a MySQL service, start it, reset its password. Copy the IP address of the master node somewhere, we will need it in step 5.
Create a PHP service, start it.
Download a WordPress tarball from http://www.wordpress.org, and expand it in your computer.
Copy file wordpress/wp-config-sample.php to wordpress/wp-config.php and edit the DB_NAME, DB_USER, DB_PASSWORD and DB_HOST variables to point to the database service. You can choose any database name for the DB_NAME variable as long as it does not contain any special character. We will reuse the same name in step 7.
Rebuild a tarball of the directory such that it will expand in the current directory rather than in a wordpress subdirectory. Upload this tarball to the PHP service, and make the new version active.
Connect to the database using the command proposed by the frontend. Create a database of the same name as in step 5 using command “CREATE DATABASE databasename;“
Open the page of the PHP service, and click “access application.” Your browser will display nothing because the application is not fully installed yet. Visit the same site at URL http://xxx.yyy.zzz.ttt/wp-admin/install.php and fill in the requested information (site name etc).
That’s it! The system works, and can be scaled up and down.

Note that, for this simple example, the “file upload” functionality of WordPress will not work if you scale the system up. This is because WordPress stores files in the local file system of the PHP server where the upload has been processed. If a subsequent request for this file is processed by another PHP server then the file will not be found. The solution to that issue consists in using the shared file-system service called XtreemFS to store the uploaded files.

The PHP Web hosting service¶

The PHP Web hosting service is dedicated to hosting Web applications written in PHP. It can also host static Web content.

Uploading application code¶

PHP applications can be uploaded as an archive or via the Git version control system.

Archives can be either in the tar, zip, gzip or bzip2 format.

Warning

the archive must expand in the current directory rather than in a subdirectory.

The service does not immediately use new applications when they are uploaded. The frontend shows the list of versions that have been uploaded; choose one version and click “set active” to activate it.

Note that the frontend only allows uploading archives smaller than a certain size. To upload large archives, you must use the command-line tools or Git.

The following example illustrates how to upload an archive to the service with id 1 using the cpsclient.py command line tool:

$ cpsclient.py upload_code 1 path/to/archive.zip

To enable Git-based code uploads you first need to upload your SSH public key. This can be done either using the command line tool:

$ cpsclient.py upload_key serviceid filename

An SSH public key can also be uploaded using the ConPaaS frontend by choosing the “checking out repository” option in the “Code management” section of your PHP service. Once the key is uploaded the frontend will show the git command to be executed in order to obtain a copy of the repository. The repository itself can then be used as usual. A new version of your application can be uploaded with git push.

user@host:~/code$ git add index.php
user@host:~/code$ git commit -am "New index.php version"
user@host:~/code$ git push origin master

Access the application¶

The frontend gives a link to the running application. This URL will remain valid as long as you do not stop the service.

Using PHP sessions¶

PHP normally stores session state in its main memory. When scaling up the PHP service, this creates problems because multiple PHP servers running in different VM instances cannot share their memory. To support PHP sessions the PHP service features a key-value store where session states can be transparently stored. To overwrite PHP session functions such that they make use of the shared key-value store, the PHP service includes a standard “phpsession.php” file at the beginning of every .php file of your application that uses sessions, i.e. in which function session_start() is encountered. This file overwrites the session handlers using the session_set_save_handler() function.

This modification is transparent to your application so no particular action is necessary to use PHP sessions in ConPaaS.

Debug mode¶

By default the PHP service does not display anything in case PHP errors occur while executing the application. This setting is useful for production, when you do not want to reveal internal information to external users. While developing an application it is however useful to let PHP display errors.

$ cpsclient.py toggle_debug serviceid

Adding and removing nodes¶

Like all ConPaaS services, the PHP service is elastic: service owner can add or remove nodes. The PHP service (like the Java service) belongs to a class of web services that deals with three types of nodes:

proxy: a node that is used as an entry point for the web application and as a load balancer
web: a node that deals with static pages only
backend: a node that deals with PHP requests only

When a proxy node receives a request, it redirects it to a web node if it is a request for a static page, or a backend node if it is a request for a PHP page.

If your PHP service has a slow response time, increase the number of backend nodes.

On command line, you can use cpsclient.py to add nodes. The add_nodes sub-command takes 4 arguments in that order: the PHP service identifier, the number of backend nodes, the number of web nodes and the number of proxy nodes to add. It also take a 5th optional argument that specify in which cloud nodes will be created. For example, adding two backend nodes to PHP service id 1:

$ cpsclient.py add_nodes 1 2 0 0

Adding one backend node and one web node in a cloud provider called mycloud:

$ cpsclient.py add_nodes 1 1 1 0 mycloud

You can also remove nodes using cpsclient.py. For example, the following command will remove one backend node:

$ cpsclient.py remove_nodes 1 1 0 0

Warning

Initially, an instance of each node is running on one single VM. Then, when adding a backend node, ConPaaS will move the backend node running on the first VM to a new VM. So, actually, it will not add a new backend node the first time. Requesting for one more backend node will create a new VM that will run an additional backend.

Autoscaling¶

One of the worries of a service owner is the trade-off between the performance of the service and the cost of running it. The service owner can add nodes to improve the performance of the service, which will have more nodes to balance the load, or remove nodes from the service to decrease the cost per hour, but increase the load per node.

Adding and removing nodes as described above is interactive: the service owner has to run a command line or push some buttons on the web frontend GUI. However, the service owner is not always watching for the performance of his Web service.

Autoscaling for the PHP service will add or remove nodes according to the load on the Web service. If the load on nodes running a Web service exceeds a given threshold and the autoscaling mechanism estimates that it will last, then the autoscaling mechanism will automatically add nodes for the service to balance the load. If the load on nodes running a Web service is low and the autoscaling mechanism estimates that it will last and that removing some nodes will not increase the load on nodes beyond the given threshold, then the autoscaling mechanism will automatically remove nodes from the service to decrease the cost per hour of the service.

Autoscaling for the PHP service will also take into account the different kind of nodes that the cloud providers propose. They usually propose small instances, middle range instances and large instances. So, the autoscaling mechanism will select different kind of nodes depending on the service owner strategy choice.

To enable autoscaling for the PHP service, run the command:

$ cpsclient.py on_autoscaling <sid> <adapt_interval> <response_time_threshold> <strategy>

where:

<sid> is the service identifier
<adapt_interval> is the time in minutes between automatic adaptation point
<response_time_threshold> is the desired response time in milliseconds
<strategy> is the policy used to select instance type when adding nodes, it must be one of:
- “low”: will always select the smallest (and cheapest) instance proposed by the cloud provider
- “medium_down”
- “medium”
- “medium_up”
- “high”

For example:

$ cpsclient.py on_autoscaling 1 5 2000 low

enables autoscaling for PHP service 1, with an adaptation every 5 minutes, a response time threshold of 2000 milliseconds (2 seconds), and using the strategy low. This means that every 5 minutes, autoscaling will determine if it will add nodes, remove nodes, or do nothing, by looking at the history of the Web service response time and comparing it to the desired 2000 milliseconds. According the specified “low” strategy, if it decides to create nodes, it will always select the smallest instance from the cloud provider.

Any time, the service owner may re-run the “on_autoscaling” command to tune autoscaling with different parameters:

$ cpsclient.py on_autoscaling 1 10 1500 low

this command updates the previous call to “on_autoscaling” and changes the adaptation interval to 10 minutes, and setting a lower threshold to 15000 milliseconds.

Autoscaling may be disabled by running command:

$ cpsclient.py off_autoscaling <sid>

The Java Web hosting service¶

The Java Web hosting service is dedicated to hosting Web applications written in Java using JSP or servlets. It can also host static Web content.

Uploading application code¶

Applications in the Java Web hosting service can be uploaded in the form of a war file or via the Git version control system. The service does not immediately use new applications when they are uploaded. The frontend shows the list of versions that have been uploaded; choose one version and click “set active” to activate it.

Note that the frontend only allows uploading archives smaller than a certain size. To upload large archives, you must use the command-line tools or Git.

The following example illustrates how to upload an archive with the cpsclient.py command line tool:

$ cpsclient.py upload_code serviceid archivename

To upload new versions of your application via Git, please refer to section Uploading application code.

Access the application¶

The frontend gives a link to the running application. This URL will remain valid as long as you do not stop the service.

The MySQL Database Service¶

The MySQL service is a true multi-master database cluster based on MySQL-5.5 and the Galera synchronous replication system. It is an easy-to-use, high-availability solution, which provides high system uptime, no data loss and scalability for future growth. It provides exactly the same look and feel as a regular MySQL database.

Summarizing, its advanced features are:

Synchronous replication
Active-active multi-master topology
Read and write to any cluster node
Automatic membership control, failed nodes drop from the cluster
Automatic node joining
True parallel replication, on row level
Both read and write scalability
Direct client connections, native MySQL look & feel

The Database Nodes and Load Balancer Nodes¶

The MySQL service offers the capability to instantiate multiple instances of database nodes, which can be used to increase the throughput and to improve features of fault tolerance through replication. The multi-master structure allows any database node to process incoming updates, because the replication system is responsible for propagating the data modifications made by each member to the rest of the group and resolving any conflicts that might arise between concurrent changes made by different members. These features can be used to increase the throughput of the cluster.

To obtain the better performances from a cluster, it is a best practice to use it in balanced fashion, so that each node has approximatively the same load of the others. To achieve this, the service allows users to allocate special load balancer nodes (glb_nodes) which implement load balancing. Load balancer nodes are designed to receive all incoming database queries and automatically schedule them between the database nodes, making sure they all process equivalent workload.

Resetting the User Password¶

When a MySQL service is started, a new user “mysqldb” is created with a randomly-generated password. To gain access to the database you must first reset this password. Click “Reset Password” in the front-end, and choose the new password.

Note that the user password is not kept by the ConPaaS frontend. If you forget the password the only thing you can do is reset the password again to a new value.

Accessing the database¶

The frontend provides the command-line to access the database cluster. Copy-paste this command in a terminal. You will be asked for the user password, after which you can use the database as you wish. Note that, in case the service has instantiated a load balancer, the command refers to the load balancer IP and its specific port, so the load balancer can receive all the queries and distributes them across the ordinary nodes. Note, again, that the mysqldb user has extended privileges. It can create new databases, new users etc.

Uploading a Database Dump¶

The ConPaaS frontend allows users to easily upload database dumps to a MySQL service. Note that this functionality is restricted to dumps of a relatively small size. To upload larger dumps you can always use the regular mysql command for this:

$ mysql mysql-ip-address -u mysqldb -p < dumpfile.sql

Performance Monitoring¶

The MySQL service interface provides a sophisticated mechanism to monitor the service. The user interface, in the frontend, shows a monitoring control, called “Performance Monitor”, that can be used to monitor a large cluster’s behaviour. It interacts with “Ganglia”, “Galera” and “MySQL” to obtain various kinds of information. Thus, “Performance Monitor” provides a solution for maintaining control and visibility of all nodes, with a monitoring dynamic data every few seconds.

It consists of three main components.

“Cluster usage” monitors the number of incoming SQL queries. This will let you know in advance about any overload of the resources. You will also be able to spot usage trends over time so as to get insights on when you need to add new nodes, serving the MySQL database.
The second control highlights the cluster’s performance, with a table detailing the load, memory usage, CPU utilization, and network traffic for each node of the cluster. Users can use these informations in order to detect problems in their applications. The table displays the resource utilization across all nodes, and highlight the parameters which suggest an abnormality. For example if CPU utilization is high, or free memory is very low this is shown clearly. This may mean that processes on this node will start to slow down, and that it may be time to add additional nodes to the cluster. On the other hand this may indicate a malfunction of the specific node.

In this latter case, in a multimaster system, it may be a good idea to kill the node and replace it with another one. The monitoring system also simplifies this kind of operations through buttons which allow to directly kill a specific node. Keep in mind, however, that high CPU utilization may not necessarily affect application performance.
“Galera Mean Misalignment” draws a real-time measure of the mean misalignment across the nodes. This information is derived by Galera metrics about the average length of the receive queue since the most recent status query. If this value is noticeably larger than zero, the nodes are likely to be overloaded, and cannot apply the writesets as quickly as they arrive, resulting in replication throttling.

The Scalarix key-value store service¶

The Scalarix service provides an in-memory key-value store. It is highly scalable and fault-tolerant. This service deviates slightly from the organization of other services in that it does not have a separate manager virtual machine instance. Scalarix is fully symmetric so any Scalarix node can act as a service manager.

Accessing the key-value store¶

Clients of the Scalarix service need the IP address of (at least) one node to connect to the service. Copy-paste the address of any of the running instances in the client. A good choice is the first instance in the list: when scaling the service up and down, other instances may be created or removed. The first instance will however remain across these reconfigurations, until the service is terminated.

Managing the key-value store¶

Scalarix provides its own Web-based interface to monitor the state and performance of the key-value store, manually add or query key-value pairs, etc. For convenience reasons the ConPaaS front-end provides a link to this interface.

The MapReduce service¶

The MapReduce service provides the well-known Apache Hadoop framework in ConPaaS. Once the MapReduce service is created and started, the front-end provides useful links to the Hadoop namenode, the job tracker, and to a graphical interface which allows to upload/download data to/from the service and issue MapReduce jobs.

Warning

This service requires virtual machines with at least 384 MB of RAM to function properly.

The TaskFarming service¶

The TaskFarming service provides a bag of tasks scheduler for ConPaaS. The user needs to provide a list of independent tasks to be executed on the cloud and a file system location where the tasks can read input data and/or write output data to it. The service first enters a sampling phase, where its agents sample the runtime of the given tasks on different cloud instances. The service then based on the sampled runtimes, provides the user with a list of schedules. Schedules are presented in a graph and the user can choose between cost/makespan of different schedules for the given set of tasks. After the choice is made, the service enters the execution phase and completes the execution of the rest of the tasks according to the user’s choice.

Preparing the ConPaaS services image¶

By default, the TaskFarming service can execute the user code that is supported by the default ConPaaS services image. If user’s tasks depend on specific libraries and/or applications that do not ship with the default ConPaaS services image, the user needs to configure the ConPaaS services image accordingly and use the customized image ID in ConPaaS configuration files.

The bag of tasks file¶

The bag of tasks file is a simple plain text file that contains the list of tasks along with their arguments to be executed. The tasks are separated by new lines. This file needs to be uploaded to the service, before the service can start sampling. Below is an example of a simple bag of tasks file containing three tasks:

/bin/sleep 1 && echo "slept for 1 seconds" >> /mnt/xtreemfs/log
/bin/sleep 2 && echo "slept for 2 seconds" >> /mnt/xtreemfs/log
/bin/sleep 3 && echo "slept for 3 seconds" >> /mnt/xtreemfs/log

The minimum number of tasks required by the service to start sampling is depending on the number of tasks itself, but a bag with more than thirty tasks is large enough.

The filesystem location¶

The TaskFarming service uses XtreemFS for data input/output. The actual task code can also reside in the XtreemFS. The user can optionally provide an XtreemFS location which is then mounted on TaskFarming agents.

The demo mode¶

With large bags of tasks and/or with long running tasks, the TaskFarming service can take a long time to execute the given bag. The service provides its users with a progress bar and reports the amount of money spent so far. The TaskFarming service also provides a “demo” mode where the users can try the service with custom bags without spending time and money.

The XtreemFS service¶

The XtreemFS service provides POSIX compatible storage for ConPaaS. Users can create volumes that can be mounted remotely or used by other ConPaaS services, or inside applications. An XtreemFS instance consists of multiple DIR, MRC and OSD servers. The OSDs contain the actual storage, while the DIR is a directory service and the MRC contains meta data. By default, one instance of each runs inside the first agent virtual machine and the service can be scaled up and down by adding and removing additional OSD nodes. The XtreemFS documentation can be found at http://xtreemfs.org/userguide.php.

SSL Certificates¶

The XtreemFS service uses SSL certificates for authorization and authentication. There are two types of certificates, user-certificates and client-certificates. Both certificates can additionally be flagged as administrator certificates which allows performing administrative file-systems tasks when using them to access XtreemFS. Certificates are only valid for the service that was used to create them. The generated certificates are in P12-format.

The difference between client- and user-certificates is how POSIX users and groups are handled when accessing volumes and their content. Client-certificates take the user and group with whom an XtreemFS command is called, or a mounted XtreemFS volume is accessed. So multiple users might share a single client-certificate. On the other hand, user-certificates contain a user and group inside the certificate. So usually, each user has her personal user-certificate. Both kinds of certificate can be used in parallel. Client-certificates are less secure, since the user and group with whom files are accessed can be arbitrarily changed if the mounting user has local superuser rights. So client-certificates should only be used in trusted environments.

Using the command line client, certificates can be created like this, where <adminflag> can be “true”, “yes”, or “1” to grant administrator rights:

$ cpsclient.py get_client_cert <service-id> <passphrase> <adminflag> <filename.p12>
$ cpsclient.py get_user_cert <service-id> <user> <group> <passphrase> <adminflag> <filename.p12>

Accessing volumes directly¶

Once a volume has been created, it can be directly mounted on a remote site by using the mount.xtreemfs command. A mounted volume can be used like any local POSIX-compatible filesystem. You need a certificate for mounting (see last section). The command looks like this, where <address> is the IP of an agent running an XtreemFS directory service (usually the first agent):

$ mount.xtreemfs <address>/<volume> <mount-point> --pkcs12-file-path <filename.p12> --pkcs12-passphrase <passphrase>

The volume can be unmounted with the following command:

$ fusermount -u <mount-point>

Please refer to the XtreemFS user guide (http://xtreemfs.org/userguide.php) for further details.

Policies¶

Different aspects of XtreemFS (e.g. replica- and OSD-selection) can be customised by setting certain policies. Those policies can be set via the ConPaaS command line client (recommended) or directly via xtfsutil (see the XtreemFS user guide). The commands are like follows, were <policy_type> is “osd_sel”, “replica_sel”, or “replication”:

$ cpsclient.py list_policies <service-id> <policy_type>
$ cpsclient.py set_policy <service-id> <policy_type> <volume> <policy> [factor]

Persistency¶

If the XtreemFS service is shut down, all its data is permanently lost. If persistency beyond the service runtime is needed, the XtreemFS service can be moved into a snapshot by using the download_manifest operation of the command line client.

Warning

This operation will automatically shut down the service and its application.

The whole application containing the service and all of its stored volumes with their data can be moved back into a running ConPaaS application by using the manifest operation.

The commands are:

$ cpsclient.py download_manifest <application-id> > <filename>
$ cpsclient.py manifest <filename>

Important notes¶

When a service is scaled down by removing OSDs, the data of those OSDs is migrated to the remaining OSDs. Always make sure there is enough free space for this operation to succeed. Otherwise you risk data loss. The download_manifest operation of the XtreemFS service will also shut the service down. This behaviour might differ from other ConPaaS services, but is necessary to avoid copying the whole filesystem (which would be a very expensive operation). This might change in future releases.

The HTC service¶

The HTC service provides a throughput-oriented scheduler for bags of tasks submitted on demand for ConPaaS. An initial bag of tasks is sampled generating a throughput = f(cost) function. The user is allowed at any point, including upon new tasks submission, to request the latest throughput = f(cost) function and insert his target throughput. After the first bag is sampled and submitted for execution the user is allowed to add tasks to the job with the corresponding identifier. The user is allowed at any point, including upon new tasks submission, to request the latest throughput = f(cost) function and adjust his target throughput. All tasks that are added are immediately submitted for execution using the latest configuration requested by the user, corresponding to the target throughput.

Available commands¶

start service_id: prompts the user to specify a mode (’real’ or ’demo’) and type (’batch’, ’online’ or ’workflow’) for the service. Starts the service under the selected context and initializes all the internal data structures for running the service.

stop service_id: stops and releases all running VMs that exist in the pool of workers regardless of the tasks running.

terminate service_id: stops and releases the manager VM along with the running algorithm and existing data structures.

create_worker service_id type count: adds count workers to the pool returns the worker_ids. The worker is added to the table. The manager starts the worker on a VM requested of the selected type.

remove_worker service_id worker_id: removes a worker from the condor pool. The worker_id is removed from the table.

create_job service_id .bot_file: creates a new job on the manager and returns a job_id. It uploads the .bot_file on the manager and assign a queue to the job which will contain the path of all .bot_files submitted to this job_id.

sample service_id job_id: samples the job on all available machine types in the cloud according to the HTC model.

throughput service_id: prompts the user to select a target throughput within [0,TMAX] and returns the cost for that throughput.

configuration service_id: prompts the user to select a target throughput within [0,TMAX] and returns the machine configuration required for that throughput. At this point the user can manually create the pool of workers using create_worker and remove_worker.

select service_id: prompts the user to select a target throughput within [0,TMAX] and creates the pool of workers needed to obtain that throughput.

submit service_id job_id: submits all the bags in this job_id for execution with the current configuration of workers.

add service_id job_id .bot_file: submits a .bot_file for execution on demand. The bag is executed with the existing configuration.

The Generic service¶

The Generic service facilitates the deployment of arbitrary server-side applications in the cloud. A Generic service may contain multiple Generic agents, each of them running an instance of the application.

The users can control the application’s life cycle by installing or removing code versions, running or interrupting the execution of the application or checking the status of each of the Generic agents. New Generic agents can be added or old ones removed at any time, based on the needs of the application. Moreover, additional storage volumes can be attached to agents if additional storage space is needed.

To package an application for the Generic service, the user has to provide simple scripts that guide the process of installing, running, scaling up and down, interrupting or removing an application to/form a Generic agent.

Agent roles¶

Generic agents assume two roles: the first agent started is always a “master” and all the other agents assume the role of regular “nodes”. This distinction is purely informational: there is no real difference between the two agent types, both run the same version of the application’s code and are treated by the ConPaaS system in exactly the same way. This distinction may be useful, however, when implementing some distributed algorithms in which one node must assume a specific role, such as leader or coordinator.

It is guaranteed that, as long as the Generic service is running, there will always be exactly one agent with the “master” role and the same agent will assume this role until the Generic service is stopped. Adding or removing nodes will only affect the number of regular nodes.

Packaging an application¶

To package an application for the Generic service, one needs to write various scripts which are automatically called inside agents whenever the corresponding events happen. The following scripts may be used:

init.sh – called whenever a new code version is activated. The script is automatically called for each agent as soon as the corresponding code version becomes active. The script should contain commands that initialize the environment and prepare it for the execution of the application. It is guaranteed that this script is is called before any other scripts in a specific code version.

notify.sh – called whenever a new agent is added or removed. The script is automatically called whenever a new agent is added and becomes active or is removed from the Generic service. The script may configure the application to take into account the addition or removal of a specific node or group of nodes. In order to retrieve the updated list of nodes along with their IP addresses, the script may check the content of the following file, which always contains the current list of nodes in JSON format: /var/cache/cpsagent/agents.json. Note that when multiple nodes are added or removed in a single operation, the script will be called only once for each of the remaining nodes.

run.sh – called whenever the user requests to start the application. The script should start executing the application and after the execution completes, it may return an error code that will be shown to the user. It is guaranteed that the init.sh script already finished execution before run.sh is called.

interrupt.sh – called whenever the user requests that the application is interrupted. The script should notify the application that the interruption was requested and allow it to gracefully terminate execution. It is guaranteed that interrupt.sh is only called when the application is actually running.

cleanup.sh – called whenever the user requests that the application’s code is removed from the agent. The script should remove any files that the application generated during execution and are not longer needed. After the script completes execution, a new version of the code may be activated and the init.sh script called again, so the agent needs to be reverted to a clean state.

To create an application’s package, all the previous scripts must be added to an archive in the tar, zip, gzip or bzip2 format. If there is no need to execute any tasks when a specific type of event happens, some of the previous scripts may be left empty or may even be missing completely from the application’s archive.

Warning

the archive must expand in the current directory rather than in a subdirectory.

The application’s binaries can be included in the archive only if they are small enough.

Warning

the archive is stored on the service manager instance and its contents are extracted in each agent’s root file system which usually has a very limited amount of free space (usually a little more than 100 MB), so application’s binaries can be included only if they are really small (a few MBs).

A better idea would be to attach an additional storage volume where the init.sh script can download the application’s binaries from an external location for each Generic agent. This will render the archive very small as it only contains a few scripts. This is the recommended approach.

Uploading the archive¶

An application’s package can be uploaded to the Generic service either as an archive or via the Git version control system. Either way, the code does not immediately become active and must be activated first.

Using the web frontend, the “Code management” section offers the possibility to upload a new archive to the Generic service. After the upload succeeds, the interface shows the list of versions that have been uploaded; choose one version and click “set active” to activate it. Note that the frontend only allows uploading archives smaller than a certain size. To upload large archives, you must use the command-line tools or Git. The web frontend also allows downloading or deleting a specific code version. Note that the active code version cannot be deleted.

Using the command-line interface, uploading and enabling a new code version is just as simple. The following example illustrates how to upload and activate an archive to the service with id 1 using the cpsclient.py command line tool:

$ cpsclient.py upload_code 1 test-code.tar.gz
Code version code-pw1LKs uploaded
$ cpsclient.py enable_code 1 code-pw1LKs
code-pw1LKs enabled
$ cpsclient.py list_uploads 1
current codeVersionId filename         description
------------------------------------------------------
      * code-pw1LKs   test-code.tar.gz
        code-default  code-default.tar Initial version

To download a specific code version, the following command may be used:

$ cpsclient.py download_code <serviceid> <code-version>

The archive will be downloaded using the original name in the current directory.

Warning

if another file with the same name is present in the current directory, it will be overwritten.

The command-line client also allows deleting a code version, with the exception of the currently active version:

$ cpsclient.py delete_code <serviceid> <code-version>

It is a good idea to delete the code versions which are not needed anymore, as all the available code versions are stored in the Generic manager’s file system, which has a very limited amount of available space. In contrast to the manager, the agents only store the active code version, which is replaced every time a new version becomes active.

Uploading the code using git¶

As an alternative to uploading the application’s package as stated above, the Generic service also supports uploading the package’s content using Git.

To enable Git-based code uploads, you first need to upload your SSH public key. This can be done either using the web frontend, in the “Code management” section, after selecting “checking out repository” or using the command-line client:

$ cpsclient.py upload_key <serviceid> <filename>

You can check that the key was successfully uploaded by listing the trusted SSH keys:

$ cpsclient.py list_keys <serviceid>

Once the key is uploaded, the following command has to be executed in order to obtain a copy of the repository:

$ git clone git@<generic-manager-ip>:code

The repository itself can then be used as usual. A new version of your application can be uploaded with git push:

$ cd code
$ git add {init,notify,run,interrupt,cleanup}.sh
$ git commit -m "New code version"
$ git push origin master

The git push command will trigger the updating of the available code versions. To activate the new code version, the same procedure as before must be followed. Note that, when using the web frontend, you may need to refresh the page in order to see the new code version.

To download a code version uploaded using Git, you must clone the repository and checkout a specific commit. The version number represents the first part of the commit hash, so you can use that as a parameter for the git checkout command:

$ cpsclient.py list_uploads 1
current codeVersionId filename            description
---------------------------------------------------------
        git-7235de9   7235de9             Git upload
      * code-default  code-default.tar    Initial version
$ git clone git@192.168.56.10:code
$ cd code
$ git checkout 7235de9

Deleting a specific code version uploaded using Git is not possible.

Managing storage volumes¶

Storage volumes of arbitrary size can be attached to any Generic agent. Note that, for some clouds such as Amazon EC2 and OpenStack, the volume size must be a multiple of 1 GB. In this case, if the requested size does not satisfy this constraint, it will be rounded up to the smallest size multiple of 1 GB that is greater than the requested size.

The attach or detach operations are permitted only if there are no scripts running inside the agents. This guarantees that a volume is never in use when it is detached.

To create and attach a storage volume using the web frontend, you must click the “+ add volume” link below the instance name of the agent that should have this volume attached to. A small form will expand where you can enter the volume name and the requested size. Note that the volume name must be unique, or else the volume will not be created. The volume is created and attached after pressing the “create volume” button. Depending on the cloud in use and the volume size, this operation may take a little while. Additional volumes can be attached later to the same agent if more storage space is needed.

The list of volumes attached to a specific agent is shown in the instance view of the agent, right under the instance name. For each volume, the name of the volume and the requested size is shown. To detach and delete a volume, you can press the red X icon after the volume’s size.

Warning

after a volume is detached, all data contained within it is lost forever.

Using the command-line client, a volume can be created and attached to a specific agent with the following command:

$ cpsclient.py create_volume <serviceid> <vol_name> <size> <agent_id>

Size must always be specified in MB. To find out the agent_id of a specific instance, you may issue the following command:

$ cpsclient.py list_nodes <serviceid>

The list of all storage volumes can be retrieved with:

$ cpsclient.py list_volumes <serviceid>

This command detaches and deletes a storage volume:

$ cpsclient.py delete_volume <serviceid> <agent_id>

Controlling the application’s life cycle¶

A newly started Generic service contains only one agent with the role “master”. As in the case of other ConPaaS services, nodes can be added to the service (or removed from the service) at any point in time.

In the web frontend, new Generic nodes can be added by entering the number of new nodes (in a small cell below the list of instances) and pressing the “submit” button. Entering a negative number of nodes will lead to the removal of the specified number of nodes.

On the command-line, nodes can be added with the following command:

$ cpsclient.py add_nodes <serviceid> <number_of_nodes>

Immediately after the new nodes are ready, the active code version is copied to the new nodes and the init.sh script is executed in each of the new nodes. All the other nodes which were already up before the execution of the command will be notified about the addition of the new nodes to the service, so notify.sh is executed in their case. The init.sh script is never executed twice for the same agent and the same code version.

Nodes can be removed with:

$ cpsclient.py remove_nodes <serviceid> <number_of_nodes>

After the command completes and the specified number of nodes are terminated, the notify.sh script is executed for all the remaining nodes to notify them of the change.

The Generic service also offers an easy way to run the application on every agent, interrupt a running application or cleanup the agents after the execution is completed.

In the web frontend, the run, interrupt and cleanup buttons are conveniently located on the top of the page, above the instances view. Pressing such a button will execute the corresponding script in all the agents. Above the buttons there is also a parameters field which allow the user to specify parameters which will be forwarded to the script during the execution.

On the command line, the following commands may be used:

$ cpsclient.py run <serviceid> [parameters]
$ cpsclient.py interrupt <serviceid> [parameters]
$ cpsclient.py cleanup <serviceid> [parameters]

The parameters are optional and, if not present, will be replaced by an empty list.

The run and cleanup commands cannot be issued if any scripts are still running inside at least one agent. In this case, if it is not desired to wait for them to complete execution, interrupt may be called first.

In turn, interrupt cannot be called if no scripts are running (there is nothing to interrupt). The interrupt command will execute the interrupt.sh script that tries to cleanly shut down the application. If the interrupt.sh completes execution and the application is still running, the application will be automatically killed. When interrupt.sh itself has to be killed, the interrupt command can be issued again. In this case, it will kill all the running scripts (including interrupt.sh). In the web frontend, this is highlighted by renaiming the interrupt button to kill.

Warning

issuing the interrupt command twice kills all the running scripts, including the child processes started by them!

Enabling a new code version is allowed only when no script from the current code version is currently running. If it is not desired to wait for them to complete execution, interrupt may be called first. When enabling a new code version, immediately after copying the new code to the agents, the new init.sh script is called.

Checking the status of the agents¶

The running status of the various scripts for each agent can easily be checked in both the web frontend and using the command-line interface.

In the web frontend, the instance view of each agent contains a table with the 5 scripts and each script’s running status, along with a led that codes the status using colors: light blue when the current version of the script was never executed, blinking green when the script is currently running and red when the script finished execution. In the latter case, hovering the mouse pointer over the led will indicate the return code in a tool-tip text.

With the command-line interface, the status of the scripts for each agent can be listed using the following command:

$ cpsclient.py get_script_status <serviceid>

The Generic service also facilitates retrieving the agent’s log file and the contents of standard output and error streams. In the web frontend, three links are present in the instance’s view of each agent. Using the command line, the logs can be retrieved with the following command:

$ cpsclient.py get_script_status <serviceid> <agent_id>

To find out the agent_id of a specific instance, you may issue the following command:

$ cpsclient.py list_nodes <serviceid>

ConPaaS in a VirtualBox Nutshell¶

ConPaaS in a Nutshell is a version of ConPaaS which runs inside a single VirtualBox VM. It is the recommended way to test the system and/or to run it in a single physical machine.

Starting the Nutshell¶

The easiest way to start the Nutshell is using VirtualBox:

If you haven’t done this already, create a host-only network on VirtualBox. To do so from the VirtualBox GUI, go to: File>Preferences>Network>Host-only Networks and click add. If you already see a host-only network (probably called vboxnet0), then you do not need to add another one.
Import the Nutshell appliance using the menu File->Import Appliance, or by simply double-clicking on the file in your file manager.
Once the Nutshell has been imported, you may adjust the amount of memory and the number of CPUs you want to dedicate to it by clicking on the Nutshell, then Settings->System->Motherboard/Processor. We recommend allocating at least 3 GB of RAM for the Nutshell to function properly.
Start the Nutshell by clicking “Start”.
Once the Nutshell is started, you can log into it. Wait a few seconds until you see a login prompt. The login credentials are:
```
Username: stack
Password: contrail
```
One important piece of information which you may want to note down is the IP address assigned to the Nutshell VM. This can be used to access the web frontend directly from your machine or to SSH into the Nutshell VM in order to execute command-line interface commands or to copy files. To find it, type the following command:
```
$ ifconfig br200
```
The IP address will appear in the second line of text.

Using the Nutshell via the graphical frontend¶

You can access the ConPaaS frontend by inserting the IP address of the Nutshell VM in your Web browser, making sure to add https:// in front of it:

https://192.168.56.xxx

Note that the frontend is accessible only from your local machine. Other machines will not be able to access it. A default user is available for you, its credentials are:

ConPaaS
Username: test
Password: password

You can now use the frontend in the same way as any ConPaaS system, creating applications, services etc. Note that the services are also only accessible from your local machine.

Note that also Horizon (the Openstack dashboard) is running on it as well. In case you are curious and you want to look inside the system, Horizon can be reached (using HTTP, not HTTPS) at the same IP address:

http://192.168.56.xxx

The credentials for Horizon are:

Openstack
Username: admin
Password: password

Using the Nutshell via the command-line interface¶

You can also use the command-line to control your Nutshell installation. You need to log in as the stack user directly in the VirtualBox window or using SSH to connect to the Nutshell VM’s IP address.

On login, both the ConPaaS and OpenStack users will already be authenticated. You should be able to execute ConPaaS commands, for example starting a helloworld service can be done with:

$ cpsclient.py create helloworld

or:

$ cps-tools service create helloworld

OpenStack commands are also available. For example:

$ nova list

lists all the active instances and:

$ cinder list

lists all the existing volumes.

The Nutshell contains a Devstack installation of Openstack, therefore different services run and log on different tabs of a screen session. In order to stop, start or consult the logs of these services, connect to the screen session by executing:

$ /opt/stack/devstack/rejoin-stack.sh

Every tab in the screen session is labeled with the name of the service it belongs to. For more information on how to navigate between tabs and scroll up and down the logs, please consult the manual page for the screen command.

Changing the IP address space used by the Nutshell¶

The Nutshell VM uses an IP address assigned by the DHCP server of the host-only network of VirtualBox. In the default settings, the DHCP server uses a range from 192.168.56.101 to 192.168.56.254. If you want to change this IP range, you can go to: File>Preferences>Network>Host-only Networks, select vboxnet0 and click the edit button and then “DHCP server”.

Note that ConPaaS services running inside the Nutshell VM also need to have IP addresses assigned. This is done using OpenStack’s floating IP mechanism. The default configuration uses an IP range from 192.168.56.10 to 192.168.56.99, which does not overlap with the default one used by the DHCP server of the host-only network in VirtualBox. If you want to modify this IP range, execute the following commands on the Nutshell as the stack user:

$ nova floating-ip-bulk-delete 192.168.56.0/25
$ nova floating-ip-bulk-create --pool public --interface br200 <new_range>

The first command removes the default IP range for floating IPs and the second adds the new range. After executing these two commands, do not forget to restart the Nutshell so the changes take effect:

$ sudo reboot

Using the Nutshell to host a publicly accessible ConPaaS installation¶

The Nutshell can also be configured to host services which are accessible from the public Internet. In this case, the floating IP pool in use by OpenStack needs to be configured with an IP range that contains public IP addresses. The procedure for using such an IP range is the same as the one described above. Care must be taken so that these public IP addresses are not in use by other machines in the network and routing for this range is correctly implemented.

If the ConPaaS frontend itself needs to be publicly accessible, the host-only network of VirtualBox can be replaced with a bridged network connected to a physical network interface that provides Internet access. Note that this bridge network must use a DHCP server that assigns a public IP address to the Nutshell or, alternatively, the Nutshell can be configured to use a static IP address (for example by editing the file /etc/network/interfaces). If the Nutshell is publicly accessible, you may want to make sure that tighter security is implemented: the default user for the ConPaaS frontend is removed and access to SSH and OpenStack dashboard is blocked.