[GRLUG] Solid State Drives

[Collin Kidder]

Adam Tauno Williams wrote:
> On Mon, 2008-12-01 at 12:14 -0500, Collin Kidder wrote:
>   
>> Adam Tauno Williams wrote:
>>     
>> ll contend that swap space is good for nothing more than a safety 
>> net. If you are using 60% of an 8GB swap space then why not just load 
>> that nearly never used data from disk when you need it? 
>>     
>
> Because that adds enormous complexity as the app doesn't know what part
> of the dataset will be required.   Load and let the kernel sort it out;
> simple and effective.    And if your paging, you are getting "data from
> disk when you need it".
>   
I suppose that's true. And, not knowing what this large dataset consists 
of, I'm in no position to argue that your way isn't the best and/or 
easiest for your given dataset.

>> You do NOT want to use a 30MB/s hard drive to supplement the storage 
>> space of a 6GB/s interface.
>>     
>
> Eh?  Theoretically swap space is approximately the same speed as reading
> from a filesystem (in reality, it's faster) and it is certainly faster
> than retrieving data over a network [potentially multiple times if you
> discard results to conserve memory... or you write them out to disk
> (same place your swap is)].  So if you have more data than you can hold
> in RAM then using swap just makes sense (within some constraints, there
> is a cut off at some point).
>   

Swap might be a bit faster than file system access but can't be faster 
than the maximum throughput of the drive(s). Also, there might be some 
way of searching for the data you need rather than loading it all. But, 
once again, I suppose that depends on the data. Some things could 
probably be easily demand loaded (such as by doing SQL queries) and some 
things might not.

> If your server or workstation's storage's maximum throughput is 30MB/s
> then its crap.  For a laptop where you've only got one lousy spindle,
> then yes, you really don't want to page (assuming you have a choice).
> But a laptop is portable, so sometimes one just has to live with it.
>   
I don't believe that even desktop harddrives can sustain better than 
30-40MB/s. Now, servers can be reading off of multiple drives with 
multiple platters and can reach much higher throughput.

>   
>>  It makes no sense. Back in the old days 
>> (like in the Windows 3.1 days) it made some sense. RAM was maybe 300MB/s 
>> and the harddrive 5MB/s. That's a 60:1 ratio. RAM was very expensive and 
>> people didn't run too many concurrent apps anyway. Nowdays the 6GB/30MB 
>> ratio is 200:1 so swapping is much more costly. 
>>     
>
> You've presented no faster solution for managing very large datasets.
>   
I suppose that I have not for a general case. I just believe that most 
of the time a more efficient solution can be reached. But I suppose that 
I must concede that there are bound to be plenty of cases where this is 
either not true or not feasible given time/budget/other constraints.