One of the related trends to cloud computing is containerized computing. One can be done without the other, but they fit nicely together.
Briefly, racks of computer equipment are packed into a specialized but otherwise standard size shipping container. The racks of computer equipment can be fully configured, interconnected, and tested as part of the manufacturing process; software and even data can be installed prior to shipping as well. The container is then loaded on a truck and shipped like any other standard intermodal shipping container. When the container reaches its destination, the container is set into position, it is connected to power, network, and cooling, and because it was fully configured prior to shipping, it is now ready use, without needing any of the unpacking, assembly, reracking, rewiring, etc. that is typically done to install computers on site.
Microsoft produced a visually informative video that illustrates the concepts around containerized computing and Microsoft’s vision for the 4th generation datacenter.
How much video storage and delivery can we fit into a container?
A standard 20 foot dry shipping container measures 20′ x 8′ x 8′6″ (lwh) with interior dimensions of approximately 19′ x 7′9″ x 7′9″ (lwh). Though overly cramped, 2 rows of 10 standard racks can fit in the interior of the 20′ container. Two racks need to be allocated to core network and management equipment, leaving 18 racks for video storage and delivery. The more common 40′ container could contain rough twice as much equipment.
Let’s look at three system configuration scenarios for a 20′ container for video.
Bronze: The first scenario uses conventional commercial-off-the-shelf server and storage equipment, as could be purchased from Dell, HP, IBM, and others. This is a generic web server configuration with some attention to video handling requirements, similar to current practice for building a large website.
Silver: The second scenario also uses conventional server and storage components, but they are specially selected and configured specifically to support video storage and delivery. Much of the software environment is carefully configured for video as well.
Gold: The third scenario pushes the envelop a little bit further, requiring some engineering and development in a few selected areas of the system that improve cost performance.
All three of these system scenarios are feasible to build now. Each scenario in turn is more challenging than the prior one, and requires increasing degrees of expertise and development effort. However, none of the systems require significantly more effort than would otherwise be required for any configuration of servers of these respective sizes. Still more could be done to push the system configuration even further (a Platinum configuration), but we will leave that perhaps to a later description.
All three of these system designs support a common set of requirements: (a) It is assumed that a system will not receive regular physical maintenance. Personnel will enter the container only on very rare occasions, if ever. Instead, if a component fails, it will simply be remotely or autonomously removed from service. (b) The systems have sufficient redundancy to tolerate a failure rate of 5% per year across all components with an operational lifespan of at least 3 years. At the end of its operational usefulness, the container will be replaced, removed, and recycled. (c) All content is delivered on demand directly to the viewer. Content is not broadcast or multicast. (d) Peak system delivery utilization (Gbs) is 75% of peak system capacity. (e) Two copies of content are maintained within the system, meaning that usable content capacity (TB) is approximately half of the raw capacity. (f) The nominal video encoding rate is 2.5Mbs, which is sufficient for a sub-HDTV picture quality when using an h.264 class codec. The encoding rate determines the hours per TB of storage and the viewers per Gbs of delivered bandwidth. (g) All components of the systems are readily available or have been announced for availability soon.
Table 1 shows how these three system scenarios compare in capability and cost.
The specially configured Silver and Gold systems provide 5-6 times better price performance than the Bronze generic web server configuration. Also note that the Silver and Bronze systems cost about the same, but the Silver system provides substantially much higher performance. The Gold system costs less than twice the Silver system, but provides slightly more than twice the performance in the same footprint.
Though the Silver and Gold systems designs handle video very well, there is nothing in their design that is only useful for video. These configurations would support high bandwidth access to any large data set. Also, though the table describes high bandwidth delivery (reading) from storage, the systems would also provide high bandwidth ingest (writing) to storage, again not limited to video.
Three design notes:
(1) All three above configurations provide both the maximum storage capacity and the maximum bandwidth delivery within their design scenarios. It is possible to shift the balance of storage and bandwidth, increasing one and decreasing the other, depending on deployment needs.
(2) All three above configurations are fully dedicated to video storage and delivery. Of course, an actual deployment might be better served with a container comprised of a mix of systems and applications. For example, in addition to video delivery, a container could contain all of the web applications and associated storage for a complete web service, including content management, title library navigation and search, viewer interaction, etc. Such a container could be configured like a YouTube in the container or a NetFlix in the container or Cable in the container. Multiple containers would be used as more service capacity is needed, and they could be geographically located where most cost effective to provide that capacity.
(3) The Silver and Gold configurations require 100-200 10Gb external Ethernet ports to connect the container to the Internet. This is a fairly large number, especially when deploying multiple containers. The use of 10-20 100GbE ports makes network connection more manageable. Of course, multiple lambdas over fiber will be used. Currently, technology is at a transition point for how many bits per lambda can be transported and how multiple lambdas are aggregated.
How much video delivery would these systems support?
Extrapolating from the Nielsen Q4 2008 A2/M2 Report and the earlier discussion, approximately 20M viewers at peak viewing times are simultaneously watching Internet video, approximately 100M people watched the last Super Bowl, approximately 228M people are watching a video screen of some form at peak viewing times. Table 2 lists how many containers of each above configuration would be needed to support these amounts of viewing. Note that these numbers assume optimal provisioning by a single broker; in practice, many more containers would be needed.
Table 2 lists the number of containers of each configuration that would be needed to support viewing for the Internet, the Super Bowl, and all viewing screens
Table 3 lists the cost of the containers needed to support viewing for the Internet, the Super Bowl, and all viewing screens.
To put these costs in some context, NBC reported that their Super Bowl 2009 revenues were $261M. All three system configurations to support the Super Bowl cost more than total Super Bowl revenues, thus installation of these systems could not be justified solely on the basis of the Super Bowl revenues. However, Super Bowl is a single 4-6 hour event; the cost of a container initially acquired for the Super Bowl can also be amortized with video programming throughout the year and with non-video applications.
The elastic characteristics of the cloud make this fluid reallocation of computing resources possible: aggregating the necessary resources for the Super Bowl, then repurposing resources or even reallocating them afterward, until the next Super Bowl or big event comes along.
The costs of Silver and Gold configurations, at 2.5x and 1.9x Super Bowl revenue respectively, are low enough that other uses of the systems throughout the year could more easily justify such an installation. The Bronze configuration, at 12.3x Super Bowl revenue, is not a viable choice, because the premium on revenue is too great to amortize effectively.
As technology improves, cost and performance will improve, making the business case easier to justify. With Moore’s Law of doubling cost performance every 2 years, since the Silver and Gold configurations offer 5x and 6x better cost performance than the Bronze configuration, the Bronze configuration would require roughly 5 years more of technology improvement to the match the cost and performance of the Silver and Gold configurations.
Of course, installations of video in the container need not be all or nothing. The on demand provisioning for events like the Super Bowl in the cloud can occur incrementally starting with partial provisioning in the cloud while most viewers still watch the game through conventional broadcast and cable delivery. Broasdcast and cable infrastruture can be containerized as well. Then over a period of a few years, provisioning can increase as usage increases up the adoption s-curve.
In summary, the advantages of containerized computing also extend to video in the container. Video compared to other applications increases the requirements for storage inside the container and for interconnect bandwidth into and out of the container. The current cost of a container that nicely supports video fits or is close to fitting within current video business models.
References
Microsoft’s vision for the 4th generation datacenter, Dec 2 2008.
http://video.msn.com/video.aspx?vid=b4d189d3-19bd-42b3-85d7-6ca46d97fe40
Wikipedia article on Intermodal Shipping Containers
http://en.wikipedia.org/wiki/Intermodal_container#Specifications
NBC press release on Super Bowl 2009 revenues
http://sev.prnewswire.com/entertainment/20090131/NY6529331012009-1.html