Chewy509
Wotty wot wot.
Hi Guys,
Just was quickly browsing the following paper called "Optimizing NAND Flash-Based SSDs via Retention Relaxation".
The basic premise is that in order to store data for a long period with NAND flash, say hold it's contents for 10yrs (being the current industry expectation or there abouts), it takes x msec (200 msec - 1800msec depending on the NAND Flash) to perform the erase/write cycle. If we are willing to lower the expected data retention period to 1 yr, or even mere days/weeks we can improve the write latency by over 2.5x with TPC-C loads or even up to 5.7x with traditional writes that are the same size as the cell. (erase/write latency being one of the issues with write performance).
Why is that interesting? In data center or SAN environments SSDs are currently being used as cache drives in front of larger banks of 15K/10K/7.2K traditional HDDs, with retention of data in the cache is only required for hours, or at most a few days... So in theory, if the SSD only writes the NAND flash in a way that the NAND cell will only retain the data for a few days, could see improvement of the caching of the storage array, and going by the paper, by up to 2.5x the performance.
For those at home or in SMB, ZFS has the ability to set cache drives as part of the zpool (these are called ZIL), so for at home installations, this could be a nice way to increase performance of your storage setup.
However, the scary thing about this idea, is that how long before someone like OCZ does similar with their consumer SSDs in order to boost write performance? (Basically after 1yr of not writing to a cell, it loses it's state, and with only a 1yr warranty, why would any company care about performance/capability beyond the warranty period).
Food for thought.
Just was quickly browsing the following paper called "Optimizing NAND Flash-Based SSDs via Retention Relaxation".
The basic premise is that in order to store data for a long period with NAND flash, say hold it's contents for 10yrs (being the current industry expectation or there abouts), it takes x msec (200 msec - 1800msec depending on the NAND Flash) to perform the erase/write cycle. If we are willing to lower the expected data retention period to 1 yr, or even mere days/weeks we can improve the write latency by over 2.5x with TPC-C loads or even up to 5.7x with traditional writes that are the same size as the cell. (erase/write latency being one of the issues with write performance).
Why is that interesting? In data center or SAN environments SSDs are currently being used as cache drives in front of larger banks of 15K/10K/7.2K traditional HDDs, with retention of data in the cache is only required for hours, or at most a few days... So in theory, if the SSD only writes the NAND flash in a way that the NAND cell will only retain the data for a few days, could see improvement of the caching of the storage array, and going by the paper, by up to 2.5x the performance.
For those at home or in SMB, ZFS has the ability to set cache drives as part of the zpool (these are called ZIL), so for at home installations, this could be a nice way to increase performance of your storage setup.
However, the scary thing about this idea, is that how long before someone like OCZ does similar with their consumer SSDs in order to boost write performance? (Basically after 1yr of not writing to a cell, it loses it's state, and with only a 1yr warranty, why would any company care about performance/capability beyond the warranty period).
Food for thought.
