Start describing billing-related stuff.

[aquarium] / doc / arch / aquarium.tex

\documentclass[preprint,10pt]{sigplanconf}
\usepackage{amsmath}
\usepackage{amssymb}
\usepackage{graphicx}
\usepackage[british]{babel}
\usepackage{url}
\usepackage{listings}
\usepackage{color}

\newcommand{\cL}{{\cal L}}
\newcommand{\TODO}{{\sl TODO \marginpar{\sl TODO}}}

\begin{document}
\conferenceinfo{ScalaDays '12}{London, UK.}
\copyrightyear{2012}
\copyrightdata{1-59593-056-6/05/0006}

\titlebanner{DRAFT---Do not distribute}



\title{Aquarium: Billing for the Cloud in the Cloud}

\authorinfo{Georgios Gousios \and Christos KK Loverdos \and Nectarios Koziris}
{GRNet SA}
{\{gousiosg,loverdos,nkoziris\}@grnet.gr}

\maketitle
\begin{abstract}
    This paper describes the architecture for the Aquarium cloud infrastructure
    software.
\end{abstract}

\category{D.3.3}{Programming Languages}{Language Constructs and Features}[Control structures]

\terms
    Object-Oriented Programming, Philosophy

\keywords
    OOP, Ontology, Programming Philosophy

\section{Introduction}
\section{Requirements}

Aquarium was designed on a clean sheet to serve a particular purpose,
namely the provision of billing services to an IaaS infrastructure,
and also be extensible to new services. In the following sections,
we briefly present the requirements that shaped Aquarium's design.

\subsection{Application Environment}
Aquarium developed as part of the Okeanos project at GRNet. The
Okeanos project is building a full stack public IaaS system for Greek
universities, and several services on top of it. Several components comprise
the Okeanos infrastructure:

\begin{description}

    \item[Synnefo] is an IaaS management console. Users can create and start
        VMs, monitor their usage, create private internal networks among VMs
        and connect to them over the web. The service backend is based on
        Google's Ganneti for VM host management and hundrends of physical
        VM container nodes.

    \item[Archipelago] is a storage service, based on the Rados
        distributed object store. It is currently under development, and the
        plan is to act as the single point of storage for VM images, shared
        volumes and user files, providing clonable snapshots and distributed
        fault tolerance.
    
    \item[Pithos] is a user oriented file storage service. Currently it its
        second incarnation, it supports content deduplication, sharing of files
        and folders and a multitude of clients.

    \item[Astakos] is an identity consolidation system that also acts as the
        entry point to the entire infrastructure. Users can login using 
        identities from multiple systems, such as the Shibboleth (SAML) 
        federation enabled across all Greek universities or their Twitter 
        accounts.

\end{description}

While all the above systems (and several prospective ones) have different 
user interfaces and provide distinct functionality in the context of
the GRnet IaaS, they all share a common notion of \emph{resources}, access
and manipulation options to which they offer to users. 

\subsection{Sharing}

The Okeanos IaaS will support the Greek higher education, an estimated
population of 100.000 students and researchers. Each member will be granted
access to a collection of resources using her institutional account. To enforce
a limit to the resources that can be acquired from the platform, each user will
have a limited amount of credits, renewable each month, which will be allowed
to spend on any resource available through the infrastructure. Resources can
also be shared; for example files on the Pithos file service or virtual machine
images on the Archipelago storage are potentially subject to concurrent usage from 
multiple users. This means that charges for the use of a single resource
may need to be distributed among several users. Also this may mean that in order
for sharing to work correctly, users may need to transfer credits among them.


\subsection{Configuration}

Billing systems are by nature open ended. As new services are deployed, new
resources appear, while others might be phased out.  Moreover, changes to
company policies may trigger changes to price lists for those resources, while
ad-hoc requests for large scale computational resources may require special
pricing policies. In order for a billing system to be able to successfully
adapt to changing requirements, it must be able to accommodate such changes
without requiring changes to the application itself. This means that all
information required for Aquarium in order to perform a billing operation,
must be provided to it externally. Moreover, to ensure high-availability,
billing configuration should be updatable while Aquarium is running, or at
least with minimal downtime, without affecting the operation of external
systems.


\subsection{Scaling}

In the context of the Okeanos system, Aquarium provides billing services on a
per user basis for all resources exposed by other systems. As such, it is in
the critical path of user requests that modify resource state; all supported
applications must query Aquarium in order to ensure that the user has enough
credits to create a new resource. This means that for a large number of users
(given previous GRNet systems usage by the Greek research community, we
estimate a concurrency level of 30.000 users), Aquarium must update and
maintain in a queryable form their credit status, 
with soft realtime guarantees. 

Being on the critical path also means that Aquarium must be highly resilient,
too. If Aquarium fails, all supported systems will also fail. Even if Aquarium
fails for a short period of time, it must not loose any billing events, as this
will allow users to use resources without paying for them. Moreover, in case of
failure, Aquarium must not corrupt any billing data under any circumstances,
while it should reach an operating state very fast after a service restart.

\section{Architecture}
\input{arch}


\section{Implementation}


\subsection{The configuration DSL}

The configuration requirements presented above were addressed by creating a new
domain specific language ({\sc dsl}), based on the YAML format.  The DSL
enables administrators to specify billable resources, billing policies and
price lists and combine them arbitrarily into agreements applicable to specific
users, user groups or the whole system. 
The DSL supports inheritance for policies, price lists and agreements and composition in the case of agreements.
It also facilitates the
definition of generic, repeatable debiting rules, which are then used by the
system to refill the user's account with credits on a periodic based.

The DSL is in itself based on five top-level entities, namely:

\begin{description}

    \item[Resources] specify the properties of resources that Aquarium knows
        about. Apart from the expected ones (name, unit etc), a resource has
        two properties that affect billing: \textsf{costpolicy} defines the
        algorithm to be used to calculate the resource usage, while the
        \textsf{complex} attribute defines whether a resource can have one or
        many instances per user.

    \item[Pricelists] assign a price tag to each resource, within a timeframe.
    
    \item[Algorithms] specify the way the billing operation is done in response
        to a billing event. The simplest (and default) way is to multiply the 
        billable quantity with the applicable price. To enable more complex billing
        scenarios, the Aquarium DSL supports a simple imperative language with
        a number of implicit variables (e.g. \texttt{price, volume, date}) 
        that enable administrators to specify, e.g. billing algorithms that
        scale with billable volume. Similarily to pricelists, algorithms
        have an applicability timeframe attached to them.

    \item[Crediplans] define a number of credits to give to users and a repetition
        period.

    \item[Agreements] assign a name to algorithm, pricelist and creditplan triplets,
        which is then assigned to each user.

\end{description}


\begin{figure}
\lstset{language=c, basicstyle=\footnotesize,
stringstyle=\ttfamily, 
flexiblecolumns=true, aboveskip=-0.9em, belowskip=0em, lineskip=0em}

\begin{lstlisting}
resources:
  - resource:
    name: bandwidthup
    unit: MB/hr
    complex: false
    costpolicy: continuous
pricelists:
  - pricelist: 
    name: default
    bandwidthup: 0.01
    effective:
      from: 0
  - pricelist: 
    name: everyTue2
    overrides: default
    bandwidthup: 0.1
    effective:
      repeat:
      - start: "00 02 * * Tue"
        end:   "00 02 * * Wed"
      from: 1326041177        //Sun, 8 Jan 2012 18:46:27 EET
algorithms:
  - algorithm:
    name: default
    bandwidthup: $price times $volume
    effective:
      from: 0
agreements:
  - agreement:
    name: scaledbandwidth
    pricelist: everyTue2
    algorithm:
      bandwidthup: |
        if $volume gt 15 then
          $volume times $price
        elsif $volume gt 15 and volume lt 30 then
          $volume times $price times 1.2
        else
          $volume times price times 1.4
        end
\end{lstlisting}

\caption{A simple billing policy definition.} 
\label{fig:dsl}
\end{figure}

In Figure~\ref{fig:dsl}, we present the definition of a simple (albeit valid) 
policy. The policy parsing is done top down, so the order of definition 
is important. The definition starts with a resource, whose name is then
re-used in order to attach a pricelist and a price calculation algorith to it.
In the case of pricelists, we present an example of \emph{temporal overloading};
the \texttt{everyTue2} pricelist overrides the default one, but only for 
all repeating time frames between every Tuesday at 02:00 and Wednesday at
02:00, starting from the timestamp indicated at the \texttt{from} field. Another
example of overloading is presented at the definition of the agreement, which
overloads the default algorithm definition using the imperative part of the
Aquarium {\sc dsl} to provide a scaling charge algorithm.

\subsection{Billing}

Commonly to most similar systems, billing in Aquarium is the application of the
provisions of a user's contract to an incoming billing event in order to
produce an entry for the user's wallet. However, in stark contrast to most
other systems, which rely on database transactions in order to securely modify
the user's balance, Aquarium performs account updates asynchronously and
concurrently for all users.

Before delving into the details of what things Aquarium computes and how these computations are performed, some terminology is needed. The following are the basic entities  involved.

\begin{description}
\item[Credits]
The analog of money. Credits are the `universal money` within Aquarium.

\item[Resource]
A billable/chargeable entity. We generally need credits to use a resource. When a resource is used,  then consume credits. Examples of resources are the `download bandwidth` and, respectively,  the 'upload bandwidth', the `disk space` and the `VM time` to name a few. Generally speaking, the ``resource'' term specifies the type.

\item[Resource instance]
A user may have several instances of a resource type. For example, regarding a ``Virtual Machine'' resource, a user may have more then one of them. They are distinguished by their unique resource instance identifier.

\item[Resource instance state]
A value which is associated with a resource instance. Usually a floating point number, as in the $10.5$ MB designation regarding a possible ``total current downloading bandwidth''.

\item[User]
An owner of resources and credits. Users are defined externally of Aquarium.

\item[Resource event]
An event that is generated by a system, which is responsible for the resource.
The resource event describes a state change for the resource. In particular, a resource event records the time when that state change occurred (`occurredMillis` attribute) and the changed value (`value` attribute).

\item[Resource event store]
A datatabase, in the general sense, where resource events are stored.

\item[Cost policy]
A cost policy refers to a resource and it is the policy used in order to charge credits for resource usage.
  
\item[User bill]
A, usually, periodic statement of how many credits the user has consumed. It may contain detailed analysis that relates consumed credits to resources.
  
\item[Billing period]
A time period at the end of which we issue a user bill.
A billing period is made of a starting date and a duration that is a multiple of a week.
A usual billing period starts on a particular month date (eg. 3rd) and lasts for a month.
Each resource type designates what happens to its accumulated value (if any) at the beginning of the billing period. Usually, at the beginning of the billing period, the accumulating amounts of resources are set accumulating amount. For example, for a monthly billing period, the total uploading bandwidth is reset to zero every month.
   
\item[User state]
The user state is made of the following distinct parts, of which the first two can be integrated to a unifying ``resource'' concept.

\begin{enumerate}
\item User credit state, that is the total credit amount for the user.

\item User resource state, which refers to the state of each resource instance that the user owns.

\item Processing state \TODO
\end{enumerate}

\item[Resource event processing]
The set of algorithmic steps by which a resource event leads to state changes of the user state.
\item[]

\item[]

\item[]

\item[]

\item[]


\end{description}

Per resource, the charging operation is affected by the cost policy and complexity
parameters. Specifically, the 3 available cost policies affect the calculation 
of the amount of resource usage to be charged as follows:

\begin{itemize}
    \item resources employing the \textsf{continuous} cost policy are charged for
        the actual resource usage through time. When a resource event arrives,
        the previous resource state between the previous charge operation and the
        current event event timestamp is charged and the resource state is then
        updated. More formally, for continuous resources, if $f(t)$ represents
        the function of resource usage through time and $p(t)$ is the function
        representing the pricelist at time $t$, 
        then the total cost up to a 
        $c(t) = \sum_{i=0}^{t} {p(t) \times \int_0^{t}{f(t)dt}}$. Most resources
        in Aquarium are continuous, for example bandwidth and disk space.

    \item resources employing the \textsf{onoff} cost policy can be in two states:
        either switched on and actively used or switched off. Therefore, the unit
        of resource usage is time and not the actual resource usage, while the
        period of charging is calculated only when the resource is switched on.
        Virtual machines are examples of resources with the \textsf{onoff} cost
        policy.

    \item resources using the \textsf{distinct} cost policy are charged
        upon usage, without time playing a role in the charge. Such resources
        are useful for one off charges, such as the allocation of
        virtual machine or the migration of a virtual machine to a less busy
        host.

\end{itemize}

Billing events are obtained through a connection to a message queue. Upon
arrival, a billing event is stored in an immutable log, and then forwarded to
the user actor's mailbox; the calculation of the actual billing entries to be
stored in the user's wallet is done within the context of the user actor,
serially for each incoming events. This permits the actor to have mutable state
internally (as described in Section~\ref{sec:ustate}), without risking the
calculation correctness. The calculation process involves steps such as
validating the resource event, resolving the current state of resource affected
by the incoming resource event, deciding the value applicable pricelist and
algorithm, generating entries for the user's wallet and updating the current
resource state for the user. A significant source of complexity in the process
is the support for temporal overriding for pricelists and algorithms: within
the timeframe between resource updates, several policies or algorithms may be
active. The billing algorithm must therefore split the billing period to pieces
according the applicability of each policy/algorithm and make sure that at 
least a baseline policy is in effect in order to perform the calculation.
Consequently, a resource event might lead to several entries to the user's wallet.

The actual format of the event is presented in Figure~\ref{fig:resevt}.

\begin{figure}
\lstset{language=C, basicstyle=\footnotesize,
stringstyle=\ttfamily, 
flexiblecolumns=true, aboveskip=-0.9em, belowskip=0em, lineskip=0em}

\begin{lstlisting}
{
  "id":"4b3288b57e5c1b08a67147c495e54a68655fdab8",
  "occured":1314829876295,
  "userId":31,
  "cliendId":3,
  "resource":"vmtime",
  "eventVersion":1,
  "value": 1,
  "details":{
    "vmid":"3300",
    "action": "on"
  }
}
\end{lstlisting}
\caption{A billing event example} 
\label{fig:resevt}

\end{figure}

\subsection{User State}
\label{sec:ustate}

\section{Performance}

To evaluate the performance and scalability of Aquarium, we performed two
experiments: The first one is a micro-benchmark that measures the time required
for the basic processing operation performed by Aquarium, which is billing for
increasing number of messages. The second one demonstrates Aquarium's
scalability on a single node with respect to the number of users.  In both
cases, Aquarium was run on a MacBookPro featuring a quad core 2.33{\sc g}hz
Intel i7 processor and 8{\sc gb} of {\sc ram}. We selected Rabbit{\sc mq} and
Mongo{\sc db} as the queue and database servers, both of which were run on a
virtualised 4 core with 4{\sc gb} {\sc ram} Debian Linux server. Both systems
were run using current versions at the time of benchmarking (2.7.1 for
Rabbit{\sc mq} and 2.6 for Mongo{\sc db}).  The two systems were connected with
a full duplex 100Mbps connection.  No particular optimization was performed on
either back-end system, nor to the {\sc jvm} that run Aquarium. 

To simulate a realistic deployment, Aquarium was configured, using the policy
{\sc dsl} to handle billing events for 4 types of resources, using 3 overloaded
pricelists, 2 overloaded algorithms, all of which were combined to 10 different
agreements, which were randomly (uniformly) assigned to users. To drive the
benchmarks, we used a synthetic load generator that worked in two stages: it
first created a configurable number of users and then produced billing events
that 


All measurements were done using the first working version of the
Aquarium deployment, so no real optimisation effort did take place. 


\section{Lessons Learned}

One of the topics of debate while designing Aquarium was the choice of
programming platform to use. With all user facing systems in the Okeanos cloud
being developed in Python and the initial Aquarium designers being beginner
Scala users (but experts in Java), the choice certainly involved risk that
management was initially reluctant to take. However, by breaking down the
requirements and considering the various safeguards that the software would
need to employ in order to satisfy them, it became clear that a
typesafe language was a hard requirement. Of the platforms examined, the {\sc
jvm} had the richest collection of ready made components; the Akka library was
particularly enticing for the scalability and distribution possibilities it
offered.

The choice of Scala at the moment it had been made was a high risk/high gain
bet for GRNet. However, the development team's experience has been generally
positive. Scala as a language was an enabling factor; case classes permitted
the expression of data models, including the configuration {\sc dsl}, that
could be easily be serialized or read back from wire formats while also
promoting immutability through the use of the \texttt{copy()} constructor. The
pervasive use of immutability allowed us to write strict, yet simple and
concise unit tests, as the number of cases to be examined was generally low.
The \textsf{Maybe}\footnote{\textsf{Maybe} works like \textsf{Option}, but it
has an extra possible state (\textsf{Failed}), which allows exceptions to be
encapsulated in the return type of a function, and then retrieved and accounted
for in a pattern matching operation with no side effects. More at
\url{https://github.com/loverdos/Maybe}} monad, enabled side-effect free
development of data processing functions, even in cases where exceptions were
the only way to go. Java interoperability was excellent, while thin Scala
wrappers around existing Java libraries enabled higher productivity and use of
Scala idioms in conjunction with Java code.

The Akka library, which is the backbone of our system, is a prime example of 
the simplicity that can be achieved by using carefully designed high-level
components. Akka's custom supervision hierarchies allowed us to partition the
system in self-healing sub-components, each of which can fail independently
of the other. For example, if the queue reader component fails due to a queue
failure, Aquarium will still be accessible and responsive for the {\sc rest}
interface. Also, Akka allowed us to easily saturate the processing components
of any system we tested Aquarium on, simply by tuning the number of threads (in
{\sc i/o} bound parts) and actors (in {\sc cpu} bound parts) per dispatcher. 

Despite the above, the experience was not as smooth as initially expected. The
most prominent problem we encountered was that of lacking documentation. The
Akka library documentation, extensive as is, only scratches the surface.
Several other libraries we use, for example Spray for {\sc rest} handling, have
non-existent documentation. The Java platform, and .Net that followed, has
shown that thorough and precise documentation are key to adoption, and we
expected a similar quality level. Related is the problem of shared community
wisdom; as most developers know, a search for any programming problem will
reveal several straightforward Java or scripting language sources. The
situation with Scala is usually the opposite; the expressive power of the
language makes it the current language of choice for treating esoteric
functional programming concepts, while simple topics are often neglected. Scala
has several libraries of algebraic datatypes but no {\sc yaml} parser. As Scala
gains mainstream adoption, we hope that such problems will fade.

From a software engineering point of view, the current state of the project was
reached using about 6 person months of effort, 2 of which were devoted to
requirements elicitation, prototype building and familiarizing with the
language. The source code currently consists of 5.000 lines of executable
statements (including about 1.000 lines of tests), divided in about 10
packages. The system is built using both {\sc sbt} and Maven. 

\section{Related Work}

\section{Conclusions and Future Work}
In this paper, we presented Aquarium, a high-performance, generic billing 
system, currently tuned for cloud applications. We presented the requirements
that underpinned its design, outlined the architectural decisions made
and analysed its implementation and performance.

Scala has been an enabling factor for the implementation of Aquarium, both
at the system prototyping phase and during actual development. 

Aquarium is still under development, with a first stable version 
being planned for early 2012. 

Aquarium is available under an open source license at 
\url{https://code.grnet.gr/projects/aquarium}.

\bibliographystyle{abbrvnat}
\bibliography{aquarium}

\end{document}