Quality of Service

As a result of emerging real-time and mission-critical applications, enterprise customers realize that the traditional "Best Effort" IP network service model is unsuitable. The main concern is that poorly behaved flows adversely affect other flows that share the same resources. It is difficult to tune resources to meet the requirements of all deployed applications.

Quality of Service (QoS) measures the ability of network and computing systems to provide different levels of services to selected applications and associated network flows. Customers that deploy mission-critical applications and real-time applications have an economic incentive to invest in QoS capabilities so that acceptable response times are guaranteed within certain tolerances.

The Need for QoS

To understand why QoS is critical, it helps to understand what has happened to enterprise applications over the past decade. In the late 1980s and early 1990s, the client/server was the dominant architecture. The main principle involved a thick client and local server, where 80 percent of the traffic was from the client to a local server and 20 percent of the client traffic needed to traverse the corporate backbone. In the late 1990s, with the rapid adoption of Internet-based applications, the architecture changed to a thin client, and servers were located anywhere and everywhere. This had one significant implication: The network became a critically shared resource, where priority traffic was dangerously impacted by nonessential traffic. A common example is the difference between downloading images and processing sales orders. Different applications have different resource needs. The following section describes why different applications have different QoS requirements and why QoS is a critical resource for enterprise data centers and service providers whose customers drive the demand for QoS.

Classes of Applications

There are five classes of applications, having different network and computing requirements. They are:

• Data transfers

• Video and voice streaming

• Interactive video and voice

• Mission-critical

• Web-based

These classes are important in classifying, prioritizing, and implementing QoS. The following sections detail these five classes.

Data Transfers

Data transfers include applications such as FTP, email, and database backup. Data transfers tend to have zero tolerances for packet loss and high tolerances for delay and jitter. Typical acceptable response times range from a few seconds for FTP transfers to hours for email. Bandwidth requirements in the order of Kbyte/sec are acceptable, depending on the file size, which keeps response times to a few seconds. Depending on the characteristics of the application (for example, the size of a file), disk I/O transfer times can contribute cumulatively to delays along with network bottlenecks.

Video and Voice Streaming

Video and voice streaming includes applications such as Apple QuickTime Streaming or Real Networks streaming video and voice products. Video and voice streams have low tolerances for packet loss, and medium tolerances for delay and jitter. Typical acceptable response times are only a few seconds. This is possible because the server can pre-buffer multimedia data on the client to a certain degree. This buffer drains at a constant rate on the client side, while simultaneously receiving bursty streaming data from the server with variations in delay. As long as the buffer can absorb all variations (without draining to empty), the client receives a constant stream of video and voice. Typical bandwidth requirements are about one Mbyte/sec, depending on the frame rate, compression/decompression algorithms, and the size of images. Disk I/O and CPU also contribute to delays. Large MPEG files must be read from disks and compression/decompression algorithms.

Interactive Video and Voice

Interactive video and voice tends to have low tolerance for packet loss and low tolerance for delay and jitter. Typical bandwidth requirements are tremendous (depending on the number of simultaneous participants in the conference, growing exponentially). Due to the interactive nature of the data being transferred, tolerances are very low for delay and jitter. As soon as one participant moves or talks, all other participants need to immediately see and hear this change. Response time requirements range from 250 to 500 milliseconds. This response time is compounded by the bandwidth requirements, with each stream requiring a few Mbit/sec. In a conference of five participants, each participant pumps out a voice and video stream while at the same time receiving streams from the other participants.

Mission-Critical Applications

Mission-critical applications vary in bandwidth requirements, but they tend to have zero tolerance for packet loss. Depending on the application, bandwidth requirements are about one Kbyte/sec. Response times range from 500 ms to a few seconds. Server resource requirements (CPU, disk, and memory) vary, depending on the application.

Web-Based Applications

Web-based applications tend to have low bandwidth requirements (unless large image files are associated with the requested Web page) and grow in CPU and disk requirements, due to dynamically generated Web pages and Web transaction-based applications. Response time requirements range from 500 milliseconds to one second.

Different classes of applications have different network and computing requirements. The challenge is to align the network and computing services to the application's service requirements from a performance perspective.

Service Requirements for Applications

The two most common approaches used to satisfy the service requirements for applications are:

• Overprovisioning

• Managing and controlling

Overprovisioning allows overallocation of resources to meet or exceed peak load requirements. Depending on the deployment, overprovisioning can be viable if it is a simple matter of just upgrading to faster lLAN switches and NICs or adding memory, CPUs, or disks. However, overprovisioning might not be viable in certain cases, for example when dealing with relatively expensive long-haul WAN links, resources that on average are underutilized, or sources that are busy only during short peak periods.

Managing and controlling allows allocation of network and computing resources. Better management of existing resources attempts to optimize utilization of existing resources such as limited bandwidth, CPU cycles, and network switch buffer memory.