Re: News : ARM Trying to Buy AmperComputing

On 22/01/2025 20:08, Rich wrote:

The Natural Philosopher <tnp@invalid.invalid> wrote:

On 21/01/2025 19:56, Rich wrote:

For SSD's, writes occur to an "erased" flash block (typically much
larger than a "disk sector" size used by the host) and given enough
writes over a short enough timeframe the SSD controller can run out of
"pre-erased" blocks to use, and when that happens write speed slows
down to the rate that can be done when a "block erase" has to occur
before the actual writes can hit the media.  Note that this "block
erase" can also invove moving any partially used data sectors out of
the block into another block, creating a "write amplification"
situation as well.

>
One of the best ways to gain speed and longevity is to buy an SSD that
is way larger than you need. So it always has empty blocks available.

 And hope it has a decent block layer that spreads the writes around so
no one part of the flash is written to more than any other parts.
 

That is pretty much standard practice these days Its called 'wear levelling'

And can do the block erases in background

 Most SSD's do erases in the background, so that they can have empty
blocks waiting to absorb writes.
 

...Exactly...

But stream a large enough set of writes at the drive, and you can (for
some drives) use up the queue of background erased blocks and then you
see your write speed drop by a good amount because the block erases are
now also part of the "write path" for your data heading at the drive.

....that was precisely the point I was making.
--
"If you don’t read the news paper, you are un-informed. If you read the news paper, you are mis-informed."
Mark Twain