[Beowulf] Re: ECC support on motherboards?

[Jim Lux]

Quoting Bill Broadley <bill at cse.ucdavis.edu>, on Thu 22 May 2008  
12:42:20 AM PDT:

> Jim Lux wrote:
>> Actually, not a big deal. The wearout is with erases/writes, not   
>> reads.  What they do is not use the same physical location for a   
>> given block.  That is, when you read/change/write a block back, it   
>> gets written to a different location.   There's a systematic way to  
>>  keep all this straight, but the net result is that writes are   
>> evenly distributed across the device.
>
> Agreed, but only when they = SSD, CF, SDHC and the like.  I believe
> this does not apply to "flash".  The flash in your PDA, motherboard,
> cell phone, mp3 player, etc does not have any innate load leveling in
> it.  Does it?

Sure it does, a lot of times.  Either integrated into the actual flash  
controller ASIC or part of the software.

On a mobo (the BIOS), you're probably right.  But how many times do  
you reflash the BIOS?  It's not a big wearout item in that  
application.  In a PDA, they almost certainly use wear leveling.


>
>> Another significant feature of Flash RAM is that writes/erases are   
>> MUCH slower than reads (even for the NAND flash that everyone uses   
>> these days) But, still, blindingly fast compared to waiting for a   
>> disk to rotate under the read/write head, so you can keep stuff in   
>> a cache, and blast it out to the disk when you shutdown, before the  
>>  power supply goes away. (Most power supplies, for instance, can   
>> hold up for 8 milliseconds.. that's an eternity when your write   
>> time is measured in nanoseconds)
>
> Er, I've heard random write times in the 1/50th to 1/10th of a second
> range, thus the reason people return expensive apple laptops with the
> $1k SSD option because outlook runs dramatically slower on their large
> mailboxes (a workloads that does lots of random tiny reads).

That would be indicative of a terrible implementation, or the use of  
NOR flash then.  Particularly for SSDs, which almost certainly use  
NAND flash, the write speed is very fast. It's the erase speed which  
is slow. Some data from Toshiba I happen to have convenient gives the  
following numbers:

NAND SLC 24 MB/s read, 8MB/s write, 2ms erase (smaller parts)
NAND MLC 18.6 MB/s read, 2.4 MB/s write, 2ms erase  (bigger density)
NOR MLC  103 MB/s read, 0.47 MB/s, 900ms erase

There's also a difference in how you access it.  NOR Flash is like  
memory (address and data lines), NAND flash is more like a I/O  
device.. you tell where to start and you get streams of data in or  
out.  The Flash used for a BIOS would be NOR flash, for this reason.

Sure, 2 MB/s isn't as fast as SDRAM, and even slower than some hard  
disk drives, but don't forget there's zero rotational latency and zero  
seek time, so overall, a flash based drive will probably be faster  
than a spinning platter drive (aside from the much lower power, etc.)   
If you were streaming video to the "disk", then you need to start  
thinking magnetic media, rather than flash.


   Amusingly
> executives with giant outlook mailboxes seem quite attracted to airbook
> SSD laptops.

Well, I just got a macbook air at work (with conventional disk) to  
replace my HP  Tablet and it *is* a nice package.  The power  
consumption of the SSD is a lot lower and it's more rugged, so if the  
executive drops their macbook air while walking up the airstair to  
their G IV, it's less likely to break.


>
> Seems common for the SSDs to be 2-10 times slower at random writes and 1/3rd
> of the bandwidth (read or write) of common cheap SATA desktop disks.

That's probably about right.

> So sure on a device that expects mostly reads (MP3 players, light
> desktop use, etc.) it's quiet, reliable, and has better battery life.
> But random writes or bandwidth aren't advantages... at least yet.
>
> For these reasons it seems that for laptop/desktop use the main
> advantages are:
> * Lower power
> * smaller size
> * higher vibration resistance
> * faster random reads (read that boot faster)
> * a low power buffer (so you spin up the disk less often)

as in "no moving parts!"  always a good thing.