Goosing flash for more efficient data centers
Data centers depend on reliable, fast caching to return database queries. Web services like Google or Facebook might have 1,000 servers dedicated to storing the results of common queries to speed up user access.
These cache servers generally rely on dynamic RAM -- it's fast, but it's expensive and power hungry. The price tag on flash memory, on the other hand, is only about one-tenth that of DRAM. It uses only about 5 percent as much energy and has about 100 times the storage density as DRAM. What's more, if the power goes out the data doesn’t disappear from flash memory. So if you run a data center, wouldn’t it make sense to lower your energy costs and boost your capacity with a move to flash memory?
The snag has always been speed. Flash memory lags far behind DRAM, particularly in write operations. Flash memory doesn’t overwrite existing data at the byte level; rather, it must write in entire blocks. That means if a block contains data, the entire block must be erased before it can be written to, which is time consuming.
Researchers at MIT, however, have created a new system for data center caching using flash memory that is competitive with existing DRAM implementations. The system, BlueCache, relies on three techniques, the first of which is actually a tried-and-true method of improving processing performance -- pipelining. While it takes a single query in flash memory approximately 200 microseconds to process, with pipelining subsequent queries are sent to the cache before the result of the first query is received.
While pipelining is a technique that has been in use for some time, MIT computer science and engineering Professor Arvind said that “we are doing it in a very deep fashion,” with more dependent steps.
The second trick to boosting flash memory performance was adding a small amount of DRAM to the system -- a few megabytes of DRAM for each million megabytes of flash. The faster-performing DRAM is used to store tables that pair data queries with flash memory addresses.
Finally, Arvind’s team developed hardware for performing read, write and delete operations in flash memory, tasks that in existing cache servers are performed in software.
“Just imagine that all these key-value stores are sitting in flash memory, and you are sending me a constant stream of these three commands,” Arvind said. “So I look at the first one and I have special-purpose hardware, which will actually go and access the flash store. As soon as I have issued that, I can instantly go and look at the second one and the next one and the next one. So all this is happening in hardware. We don’t talk to the processor while we are doing this.”
Those operations, he said, can be performed much faster in hardware than in software.
"The flash-based KV store architecture developed by Arvind and his MIT team resolves many of the issues that limit the ability of today's enterprise systems to harness the full potential of flash,” Vijay Balakrishnan, director of the Data Center Performance and Ecosystem program at Samsung Semiconductor’s Memory Solutions Lab, told MIT News. "The viability of this type of system extends beyond caching, since many data-intensive applications use a KV-based software stack, which the MIT team has proven can now be eliminated."
When asked when we might see BlueCache in the market, Arvind demurred. “Many people are saying, ‘Why don’t you guys start a company?’ he said. “But really, industry has already picked up the idea, so it is happening.”
Posted by Patrick Marshall on Sep 13, 2017 at 1:49 PM