Facebook is moving to an all-Flash solid state drive environment for its data bases.
It was the first company to deploy Flash at such a large scale, using it alongside HDD and now, using new technology made available under a new Fusion-io product, announced at the Open Compute Summit today, it is making another giant leap to keep up with rising data amounts.
Facebook unveiled only yesterday a new search feature that will allow users to sift through posts, messages and pictures on the social networking site – a function that will ultimately require the fast retrieval of data.
Storage was a major part of the opening summit keynote by Frank Frankovsky, director of hardware design and supply chain at Facebook and one of the leading figures in the Open Compute Project, an open-source hardware and data center design community Facebook started in 2011.
Frankovsky showed an ARM-based storage-server motherboard by Calxeda and an Intel motherboard, both developed for OCP’s storage-chassis design Open Vault.
Fusion-io CEO David Flynn said Facebook, which previously worked in a hybrid environment using servers containing hard disk drives and Flash, is moving all its archival data – “even slow data” – onto all-Flash servers in a move to increase efficiency and performance.
He said on top of more efficient data retrieval, Facebook is likely to see a substantial reduction in its cost of power and cooling by using the ioScale product set designed with input from Facebook (which you can read more about here).
“We are talking about whether the server itself is the failure unit in the case of Facebook,” Flynn said.
“For operations like this, you don’t replace a drive. You just turn the thing off, and that servicing model is one where you will want to make that server super simple, so that dozens of components don’t need servicing.”
The ioScale Flash server is just one card, which is where its simplicity lies. Fusion-io has also made the price point lower for ioScale, with a list price of US$3.89 per Gigabyte for a 100-unit order quantity, meaning smaller operators can also make use of the technology.
“And if you put in HDDs (hard disc drives), it can take ten of them to meet the capacity point of Flash,” Flynn said.
This makes Flash much easier to scale out and Facebook, according to Flynn, wants to minimize its number of controller chips and power consumption and servers, so it doesn’t have as many things sitting in the data center that can fail.
“It ends up being cheaper for them to just turn off the server and flip over to another – they actually throw these things in a wood chipper, at least this is what was observed from Facebook’s old data center decommissioning,” Flynn said.
Flynn said in some cases, Facebook has been known to take the Fusion-io cards out of the server and placing them in new ones, a sign that Flash has become more resilient than originally thought.
“The irony here is that four years ago the entire industry – Facebook included – was saying Flash is not going to last,” Flynn said.
“It is super exciting to see that it is much more reliable and more enduring than frankly we even anticipated. Facebook found our cards were more reliable than even the RAID controller.”
Photos: Facebook engineering’s biggest headache
One of the biggest challenges of the Facebook infrastructure team is keeping up with the growth of the volume of user photos they need to store. Jay Parikh, Facebook’s VP of engineering, said in his summit keynote that the company’s data centers today are storing more than 240bn photos, with users adding more than 350m every day.
Not only do those photos need to be stored, they need to be accessible by users quickly. Not all photos, however, are accessed with the same frequency.
Facebook has broken a photo’s lifespan into three phases: hot, warm and cold, in the descending order of user activity around the photo. A photo is stored on a different storage system in each phase.
To keep a smooth user experience with photos, each storage system is optimized to deliver the best and most cost efficient performance based on the frequency of its access by users.