Turbo Charging Content Delivery Networks with Flash

by Josh Miner, Director of Product Marketing, Fusion-io

A
Frost & Sullivan report recently projected growth in video delivery alone to
reach


$1 billion dollars by 2013.
Consumers all over the world are seeking the same content simultaneously from a
host of devices. Content Delivery Networks (CDNs) are an essential tool for
delivering this information as efficiently as possible. Most networks
geographically position servers to host relevant content near requestors to
minimize latency as well as to distribute bandwidth needs across multiple sites.
Because CDNs were created to improve the user experience, they are a natural fit
for flash memory, if implemented the right way.

This article discusses some ways CDNs are leveraging flash to
address key architectural challenges to enhance performance, while at the same
time improving their business model. It includes the following topics:

• 
  Streaming more content, faster
• 
  Cooling cache pool costs
• 
  Faster, smarter content serving

Streaming More Content, Faster

CDNs typically cache as much content as they can on pooled RAM in
large server farms. Applications dynamically shuffle “hot” content from disk to
RAM and move “cooler” content to disks. As with any cache, the more data you can
put in it, the better your service levels.

The problem is that the cost both to purchase the RAM and to
maintain continually growing infrastructure, grows at a faster pace than the
company’s user-base and content library. Consequently, CDNs must limit the
amount of content they keep in RAM—and cache misses can cripple performance if,
as is often the case, some piece of cooler content unexpectedly goes viral.

Flash enables much more data to be stored in each server so that
CDNs can ensure high service levels as user-bases and content libraries grow.
Low latency flash devices have an added benefit of minimizing queuing—making
content servers much more resilient to traffic spikes.

“Our total frequently-accessed music cache now holds 10 times the
songs it used to, which both enhances existing user experience and gives us
plenty of headroom for future growth.”  

Aaron Porter, Director of System Administration, Pandora
Pandora, the leading personalized Internet radio station, added
flash to its caching servers to achieve impressive results. Aaron Porter,
Director of System Administration for Pandora said, “The [flash devices] perform
as well as our RAM caches, but offer 10 times the capacity per server. Our total
frequently-accessed music cache now holds 10 times the songs it used to, which
both enhances existing user experience and gives us plenty of headroom for
future growth.”

Similarly, Dwango, who some call the “YouTube” of Japan, added
flash to its video-sharing caching servers. Tetsuya Sato, Infrastructure Manager
for Dwango’s Nico-Nico Douga said, “Hot videos can be viewed by tens of
thousands of simultaneous users. The [flash devices] have allowed us to support
these traffic spikes with just four to five servers at the most. This allows us
to guarantee our customers a predictably high experience for all content,
regardless of popularity, and without the over-provisioning we used to need.”

Cooling Cache Pool Costs

As just mentioned, the cost to scale and maintain RAM cache pools
is a primary constraint for CDNs. If, like Pandora, you can store 10 times the
songs per server, you can maintain existing service levels over the same data,
on 1/10th
the servers. CDNs typically strike a balance between adding capacity for more
cache and future cache growth, while consolidating existing infrastructure. 
In
fact, many CDNs are finding that they can do without as many RAM-heavy servers
and rely primarily on flash for content hosting.

Pandora’s Aaron Porter said Pandora cut its existing cache
footprint by 40%. On top of this, it greatly reduced future scale out costs.
First, it extended the life of its existing storage. Aaron said, “We have
stopped purchasing SAS spindles for content delivery, which has extended the
usable life of our disk-based systems by months, if not years,” Aaron said. “In
the first quarter that we’ve run them, the [flash devices] allowed us to delay
purchasing 50 disk-based content servers—at least 1,000 SAS spindles.”

New servers will also cost less and require less RAM—and RAM’s
associated power and cooling costs. Aaron added, “The [new] servers provide cache
with order-of-magnitude higher capacities of NAND flash instead of RAM. This
allows us to continue to support our growing listener usage on fewer, smaller,
and less power-hungry servers.“

Like Pandora, video-sharing provider Dwango reduced its
caching server footprint, “We consolidated the video-hosting caching servers
that use ioDrives by a factor of four to one, cutting 200 servers to about 50
servers,” Tetsuya Sato said. “This consolidation factor fully utilizes the 10GbE
bandwidth in each server and leaves plenty of I/O headroom for complex,
on-demand traffic for a 200,000-user workload.” An additional benefit Dwango
gained—no more performance guessing games. Tetsuya said, “We also deployed a
building-block architecture so we can easily determine the number of servers we
need to meet additional workload requirements. And because we can continue to
use our favorite, off-the-shelf servers, we also ensure easy procurement.”

Faster, Smarter Content Serving

CDNs that deliver advertising or recommendations for content use
analytics, oftentimes on Big Data, to target individual users in real time.
Faster analysis means less delay and the ability to analyze more data,
increasing the value of these services.

Kontera
delivers real-time, Web-based, content-relevant advertising from offices in the
US, UK and Israel. It uses a 360GB MongoDB database to analyze the 100 million
daily page views generated by its 15,000 publisher clients, and then insert ads
relevant to page content. The aim is zero delay in the delivery of Web pages.

Kontera has eliminated any delay that its analysis might cause by
hosting its MongoDB database on flash. Greg Pendler, Kontera Production
Operations Manager, said, “Our page analysis…is a write-heavy process that must
be done in real-time. We tried using a SAN storage target, but it would have
taken 55 spindles to get us the performance we needed. We tried a local disk
array, and that still wasn’t fast enough. [Flash memory provides] all the
performance we need, and give us plenty of room to grow. Queries that used to
take minutes now complete in seconds.”

Aggregate Knowledge’s patent-pending AK
Media Intelligence Platform combines
campaign analytics with audience data mining and multi-touch attribution to make
media accountable, and to help advertisers, agencies, DSPs, ad networks, SSPs,
and publishers manage the software they need to run their businesses themselves.
Aggregate Knowledge’s Director of Operations said, “With MongoDB [using flash
memory,] we enabled sub-millisecond search times on a billion documents.”

Summary

Flash memory is an enabling technology in the growing CDN market.
It is ideal for caching, overcoming the in-server density limitations and
power-hungry cost constraints of scale-out RAM pools. It also speeds analytics
processing, enabling CDNs to ensure the content it serves is the most relevant
it can be.

About
the Author

Fusion-io has pioneered a
next generation storage memory platform for shared data decentralization
that significantly improves the processing capabilities within a
datacenter by relocating process-critical, or “active”, data from
centralized storage to the server where it is being processed, a
methodology referred to as data decentralization. Fusion’s integrated
hardware and software solutions leverage non-volatile memory to
significantly increase datacenter efficiency and offers enterprise grade
performance, reliability, availability and manageability. Fusion’s data
decentralization platform can transform legacy architectures into next
generation datacenters and allows enterprises to consolidate or
significantly reduce complex and expensive high performance storage,
high performance networking and memory-rich servers. Fusion’s platform
enables enterprises to increase the utilization, performance and
efficiency of their datacenter resources and extract greater value from
their information assets.