Joseph Loo wrote: > Bzzzz wrote: > > Andrew McGlashan wrote: > > ... > ... > > I think you are missing the problem associated with SSd. The wear problem is > associated with the amount of free space. If the drive is 99.99% full, you > could probably wear the drive out in no time at all. The wear problem is > prevented by using free space that has not been written to it. thus if 00 % > full drive will wear out faster than a drive that is 1% full. If you do not > speak in context of the percentage full the benchmarks are not too useful.
That depends upon the quality of the device. Higher quality SSDs always have internal space in which to move data around. This quality issue was especially notable with earlier devices and with bad clones. A lot of early nand flash devices were terrible! But now high quality vendors always include unadvertised internal space in which to operate for wear leveling. Also remember that SSDs operate acceptably well without trim telling it that blocks are free. Sure trim can help improve things! I am not discounting trim. But SSDs still operate even when used in a system that does not have knowledge of trim. How many people have installed an SSD in their laptop or other mobile device as a spinning disk replacement but never enabled trim on that device? (Me for one.) In that case the drive does not have knowlege of file system structure. In that case the drive always thinks it is 100% full. It never gets a trim erase instruction. It only gets write instructions. I will take a guess that most SSDs in actual operation today always think they are 100% full. That will change as more systems with SSDs come into the world that have the trim feature operating. But even with trim letting the drive know that some blocks are available for erase if you fill the disk up to 90% and then don't rewrite any of those files again then you *think* those files are never (re)written again. But that isn't the way SSD firmware works. SSD firmware reserves free space internally to the drive and mixes it into the pool. For wear leveling the firmware will copy data from one block to another. It won't simply hammer away at a single block re-writing it endlessly. It really can't since it takes so long to erase. It must pre-erase blocks that it intends to use later. Internally those will be erased in large chunks. Therefore the firmware will always arrange to pre-erase as much as possible. The trim command helps but with reserved space it is a help not a requirement. > Most consumer grade ssd are limited to about 10K writes per cell. If you > exceed the limit, dead cell. Remember that another factor involve is the > number of spare cells. all of these things play in the role when an ssd > fails. Say a 60G device might have 64G of actual space but never exposes more than 60G to the consumer. The firmware will actually have 64G of total space to operate. It can keep 4G pre-erased in the internal free lists. The firmware sees a request to write a few blocks. It pulls some blocks off the internal free list and writes them. But while you can request a write to 512 bytes you can't write that small of a block internally in the SSD. In order to write a few bytes the SSD firmware will really need to read a much larger internal sized block, modify just those small few bytes, then write the entire larger sized block. These larger internal blocks are usually at least a megabyte. This write must be to a different area in the pool. The previous area in the pool is then available for erasing. It will be queued and erased in the background. This generally causes a migration of data across the pool. This is good for wear leveling. But writing small blocks is inefficient and therefore for performance reasons the firmware will buffer up as much as possible and try to write only larger blocks at a time. This action of needing to erase large blocks and rewriting blocks causes all of the data in the SSD to be continously re-written as the SSD is used. Which is a good thing for reliability and for wear leveling. If a block can't be written it can be marked failed internally and another block used. The performance of doing this buffer, write, erase, write is strongly influenced by the firmware. This is why SSD companies closely guard their firmware and why some companies benchmark better than others even when the underlying nand flash technology seems very similar. The controller hardware and firmware for managing it is critical to reliability and performance. It is this continous internal movement of data in an SSD which enables it to continue to function successfully even without trim operating. And it provides wear leveling. It doesn't really matter if the file system structure layer waaaay out on the consumer end is full or not. (Although having trim helps and improves things by allowing the firmware to queue those blocks for pre-erase sooner.) Larger capacity SSDs should last longer due to being able to wear level over their full capacity. A 600G SSD should last through many more consumer write cycles than a 30G SSD as it has more total blocks to wear level across. Bob
signature.asc
Description: Digital signature
