SSDs and Virtual Memory

[Collider]

In a multiple drive scenario, what would be the optimal virtual memory configuration performance-wise?

Let's say if a system has 1 or 2 SSDs and additional mechanical drives drives - just thinking about speed what would offer the best in windows performance in terms of setting up virtual drive differently from window's default?

 

[LokutusofBorg]

Unless you're cramped on space, or you have a *very* fast dedicated drive for your page file, you're better off just leaving it at Windows defaults. The IO profile of the Windows page file is one of the best suited workloads for an SSD. Just install Windows and forget about it.

 

[Collider]

Unless you're cramped on space, or you have a *very* fast dedicated drive for your page file, you're better off just leaving it at Windows defaults. The IO profile of the Windows page file is one of the best suited workloads for an SSD. Just install Windows and forget about it.

Update: I ran PCMarks to test both configurations and here's what I found:

1) Default windows settings - 10,439 PCMarks
OS/Apps/Virtual disk on 80 GB X25-M

2) Virtual drive set to a 2nd drive - 11,552 PCMarks
OS/Apps on 80 GB X25-M
Virtual disk on 2 x WD Black 640GB (RAID-0)

Initially I had the same opinion as you, but the benchmarks show 1,100 points improved score when virtual disk is setup on a separate drive.

 

[Nothinman]

Update: I ran PCMarks to test both configurations and here's what I found:

1) Default windows settings - 10,439 PCMarks
OS/Apps/Virtual disk on 80 GB X25-M

2) Virtual drive set to a 2nd drive - 11,552 PCMarks
OS/Apps on 80 GB X25-M
Virtual disk on 2 x WD Black 640GB (RAID-0)

Initially I had the same opinion as you, but the benchmarks show 1,100 points improved score when virtual disk is setup on a separate drive.

2 runs isn't nearly enough for meaningful results and I don't think "PCMarks" actually equates to anything in the real world anyway.

 

[Emulex]

its rather simple - you have multiple threads to dispatch i/o; one might suspect if you have a swap on both drives it could increase performance more. the more i/o your system can dispatch (here is where more drives, more cores, o/s choice) can really shine.

In the old days folks used many files on many luns for sql to increase i/o - you can still do this but as microsoft and storage (tiering) becomes better you may not have to physically place them on separate locations.


to keep it simple though: even though a raid-0 could be 2x faster - it presents as one drive and may dispatch i/o with one thread - where as the same drives separate may be capable of more i/o depending on the application working separately since you can dispatch two worker threads to do i/o.

(the x25-m G1/G2 isn't the fastest drive at write speeds since it doesn't use write-back caching which is dangerous; most hard drives and sandforce/indilinx ssd use dangerous write-back caching without batteries currently - to me i'd rather play it safe than risk catastrophic data corruption).

 

[Dadofamunky]

its rather simple - you have multiple threads to dispatch i/o; one might suspect if you have a swap on both drives it could increase performance more. the more i/o your system can dispatch (here is where more drives, more cores, o/s choice) can really shine.

In the old days folks used many files on many luns for sql to increase i/o - you can still do this but as microsoft and storage (tiering) becomes better you may not have to physically place them on separate locations.


to keep it simple though: even though a raid-0 could be 2x faster - it presents as one drive and may dispatch i/o with one thread - where as the same drives separate may be capable of more i/o depending on the application working separately since you can dispatch two worker threads to do i/o.

(the x25-m G1/G2 isn't the fastest drive at write speeds since it doesn't use write-back caching which is dangerous; most hard drives and sandforce/indilinx ssd use dangerous write-back caching without batteries currently - to me i'd rather play it safe than risk catastrophic data corruption).

So Sandforce and Indilinx aren't crash-consistent? That's not good. I'd prefer to sacrifice a little speed for safety.