Second day of the 30th IEEE International Conference on Massive Storage Systems and Technologies. The morning led off with Russ McElroy from eBay outlining their IT goals and directions, title of his talk was Scalable Digital Data Center Design. Russ runs all of storage for eBay and Paypal. Principles driving eBay's storage system designs include:
- software defined everything
- 2 levels of FLASH: fast and "big data" FLASH
- hyperscale Openstack
eBay has a tight cost model that tracks capes and opex for proposed IT investments. IT opex is a big part of their costs.
Russ talked about a shifting paradigm of storage: stated that disk is not keeping up in terms of areal density. He also sees commoditization of FLASH arrays and hyper scale vendors like eBay being able to develop a software defined storage ecosystem and build their own ODM-based solutions using dense storage enclosures. Combined with RDMA/IP at 40 Gbps with ethernet connectivity for storage, givens them low cost and performance. He said that storage companies need to adapt to this new model. So with 4TB flash drive, can get 100-200 TB per rack unit for what eBay can design themselves.
FLASH is FLASH, not disk. Software question: how do you use FLASH? Optimizations needed in: kernels, applications, controllers, file systems, software defined storage (design with FLASH in mind). All these things need to change to adapt to FLASH.
Russ stated big data FLASH is coming. Design criteria: just has to beat disk in IOPS per TB; heavy read workload, write seldom; 30c/GB target; 30 day power-off retention and < 200ms power on response; 6 GB/s throughput; 8 PB in rack. Hadoop requirements: tiered replicas; abstract storage including data nodes separated from Map-Reducers; line rate networking.
Storage Class Memory: Russ talked about replacing memory with persistent technologies, including PCM, restive RAM, and Spintronics. eBay likes Memrister as a potential FLASH replacement going forward.
Challenges for their plan:
- how to control costs to use FLASH for big data
- software not understanding FLASH
- storage systems not understanding FLASH
- CAPEX gap makes adoption more difficult
- when will FLASH pass the baton