Scalability is the capability of a system, network, or process to handle a growing amount of work, or its potential to be enlarged in order to accommodate that growth. For example, it can refer to the capability of a system to increase its total output under an increased load when resources (typically hardware) are added. An analogous meaning is implied when the word is used in an economic context, where scalability of a company implies that the underlying business model offers the potential for economic growth within the company.
Contents
- Measures
- Examples
- Horizontal and vertical scaling
- Database scalability
- Strong versus eventual consistency storage
- Performance tuning versus hardware scalability
- Weak versus strong scaling
- References
Scalability, as a property of systems, is generally difficult to define and in any particular case it is necessary to define the specific requirements for scalability on those dimensions that are deemed important. It is a highly significant issue in electronics systems, databases, routers, and networking. A system whose performance improves after adding hardware, proportionally to the capacity added, is said to be a scalable system.
An algorithm, design, networking protocol, program, or other system is said to scale if it is suitably efficient and practical when applied to large situations (e.g. a large input data set, a large number of outputs or users, or a large number of participating nodes in the case of a distributed system). If the design or system fails when a quantity increases, it does not scale. In practice, if there are a large number of things (n) that affect scaling, then resource requirements (for example, algorithmic time-complexity) must grow less than n2 as n increases. An example is a search engine, which scales not only for the number of users, but also for the number of objects it indexes. Scalability refers to the ability of a site to increase in size as demand warrants.
The concept of scalability is desirable in technology as well as business settings. The base concept is consistent – the ability for a business or technology to accept increased volume without impacting the contribution margin (= revenue − variable costs). For example, a given piece of equipment may have a capacity for 1–1000 users, while beyond 1000 users additional equipment is needed or performance will decline (variable costs will increase and reduce contribution margin).
Measures
Scalability can be measured in various dimensions, such as:
Examples
Horizontal and vertical scaling
Methods of adding more resources for a particular application fall into two broad categories: horizontal and vertical scaling.
There are tradeoffs between the two models. Larger numbers of computers means increased management complexity, as well as a more complex programming model and issues such as throughput and latency between nodes; also, some applications do not lend themselves to a distributed computing model. In the past, the price difference between the two models has favored "scale up" computing for those applications that fit its paradigm, but recent advances in virtualization technology have blurred that advantage, since deploying a new virtual system over a hypervisor (where possible) is often less expensive than actually buying and installing a real one. Configuring an existing idle system has always been less expensive than buying, installing, and configuring a new one, regardless of the model.
Note, that NFV defines these terms differently: scaling out/in is the ability to scale by add/remove resource instances (e.g. virtual machine), whereas scaling up/down is the ability to scale by changing allocated resources (e.g. memory/CPU/storage capacity)
Database scalability
A number of different approaches enable databases to grow to very large size while supporting an ever-increasing rate of transactions per second. Not to be discounted, of course, is the rapid pace of hardware advances in both the speed and capacity of mass storage devices, as well as similar advances in CPU and networking speed.
One technique supported by most of the major database management system (DBMS) products is the partitioning of large tables, based on ranges of values in a key field. In this manner, the database can be scaled out across a cluster of separate database servers. Also, with the advent of 64-bit microprocessors, multi-core CPUs, and large SMP multiprocessors, DBMS vendors have been at the forefront of supporting multi-threaded implementations that substantially scale up transaction processing capacity.
Network-attached storage (NAS) and Storage area networks (SANs) coupled with fast local area networks and Fibre Channel technology enable still larger, more loosely coupled configurations of databases and distributed computing power. The widely supported X/Open XA standard employs a global transaction monitor to coordinate distributed transactions among semi-autonomous XA-compliant database resources. Oracle RAC uses a different model to achieve scalability, based on a "shared-everything" architecture that relies upon high-speed connections between servers.
While DBMS vendors debate the relative merits of their favored designs, some companies and researchers question the inherent limitations of relational database management systems. GigaSpaces, for example, contends that an entirely different model of distributed data access and transaction processing, space-based architecture, is required to achieve the highest performance and scalability. On the other hand, Base One makes the case for extreme scalability without departing from mainstream relational database technology. For specialized applications, NoSQL architectures such as Google's BigTable can further enhance scalability. Google's massively distributed Spanner technology, positioned as a successor to BigTable, supports general-purpose database transactions and provides a more conventional SQL-based query language.
Strong versus eventual consistency (storage)
In the context of scale-out data storage, scalability is defined as the maximum storage cluster size which guarantees full data consistency, meaning there is only ever one valid version of stored data in the whole cluster, independently from the number of redundant physical data copies. Clusters which provide "lazy" redundancy by updating copies in an asynchronous fashion are called 'eventually consistent'. This type of scale-out design is suitable when availability and responsiveness are rated higher than consistency, which is true for many web file hosting services or web caches (if you want the latest version, wait some seconds for it to propagate). For all classical transaction-oriented applications, this design should be avoided.
Many open source and even commercial scale-out storage clusters, especially those built on top of standard PC hardware and networks, provide eventual consistency only. Idem some NoSQL databases like CouchDB and others mentioned above. Write operations invalidate other copies, but often don't wait for their acknowledgements. Read operations typically don't check every redundant copy prior to answering, potentially missing the preceding write operation. The large amount of metadata signal traffic would require specialized hardware and short distances to be handled with acceptable performance (i.e. act like a non-clustered storage device or database).
Whenever strong data consistency is expected, look for these indicators:
Indicators for eventually consistent designs (not suitable for transactional applications!) are:
Performance tuning versus hardware scalability
It is often advised to focus system design on hardware scalability rather than on capacity. It is typically cheaper to add a new node to a system in order to achieve improved performance than to partake in performance tuning to improve the capacity that each node can handle. But this approach can have diminishing returns (as discussed in performance engineering). For example: suppose 70% of a program can be sped up if parallelized and run on multiple CPUs instead of one. If
Substituting the value for this example, using 4 processors we get
If we double the compute power to 8 processors we get
Doubling the processing power has only improved the speedup by roughly one-fifth. If the whole problem was parallelizable, we would, of course, expect the speed up to double also. Therefore, throwing in more hardware is not necessarily the optimal approach.
Weak versus strong scaling
In the context of high performance computing there are two common notions of scalability: