Linus Torvalds (torvalds@linux-foundation.org) on 1/22/11 wrote:
---------------------------
>iz (i@z.x) on 1/22/11 wrote:
>>
>>If space is more important than performance, use a low reserve percentage (the
>>standard 7% from power-of-two to HD power of 10 trick + round down to nearest ten
>>GB: e.g. 128GB real -> 120GB marketing is IMHO the minimum).
>
>Hmm. Intel seems to be doing on with just the 7% difference
>between GB and GiB. Afaik, those 160GB drives are 160GiB,
>there's no extra rounding down.

That is because if you don't write a lot fast enough to the
disk, it has enough time to do the GC in the background.
Which is as it should be, and not a big problem because the
max write speed is way faster than what the disk can handle
endurance wise.

Another factor is that the max slow down factor of
100/7 = 14 is the worst case, and there are always blocks
that are less fragmented than that. So usually little or
not GC needs to happen to get a free block. How much the
average block free percentage is depends on the workload,
but should be much higher than the worst case 7%.

The hidden cost is of course less endurance and extra
power usage because more writes are done and the effective
write amplification is higher.

>But I have this dim memory of (I think) Hitachi selling an
>"enterprise" version of the Intel SSD where the difference
>was literally just the extra reserve space for added
>performance and lifetime.
>
>But maybe it's my meds talking.

Any idea what their reserve ratio is for their enterprise
disks? Or is SLC the only difference?

>
>>NAND flash chips more or less come in multiples of 8GB, so
>>you can usually guess what reserve ratio a disk has. If it's sold as a "256GB"
>>disk I wouldn't buy it because performance is clearly not a priority. It should be sold as 250 at most.
>
>I suspect how much reserve space is more an issue of how
>good your GC/writeleveling algorithm is.

It primarily depends on the IO pattern.

The max fragmentation slowdown is defined by the reserve ratio only.

The extra slowdown of doing suboptimal GC/wearleveling is
a constant factor (or it's so crap it collapes badly when
it can't easily find easy to reclaim blocks).

FTL crappiness can be partly hidden with higher reserve ratios. But then they want you to really use TRIM as well.

Example for a disk with 7% reserve and 200MB/s max write
throughput. Min write speed will be 200/(100/7) = 14 MB/s.
If the disk had 14% reserve it would be 28 MB/s. So the
worst case speed is twice as high for only 7% extra reserve.
It's hard to tell how much it will be in practise, but 25%
to 50% seem likely.

> And how aggressive you are.

Having lots of erased blocks ready to use speeds up write
bursts, but the more you pre-erase, the more unnecessary GC
is done. This is a lot easier for bigger disks because
they can afford to pre-erase more without worrying about destroying GC efficiency.

> The 7% thing seems to work for Intel, but they
>have slower top write throughput than some of the new
>controllers. The two may be related.

Probably, but the newer disks may also use newer, faster
flash.