SSD Speed Limited by CPU?

[ksec]

There were tests that show over-clocking an CPU will have increased transfer speed with Vertex 3 Pro.

Any technical reason behind this?

 

[Burner27]

Link please.

 

[Nothinman]

I would think they'd be more limited by the SATA interface. If that's really the case then I would think that would point to a problem with either the firmware on that drive or the drivers for the SATA controller.

 

[Rubycon]

Overclocking PCI-E speed can increase bandwidth but can also corrupt your data too!

If a benchmark is influenced by operating system caching schemas then certainly those numbers will reflect CPU/memory overclocking.

 

[razel]

Depends how much % difference was the overclock and what else (like PCI-E speed) was affected by the overclock. I get a feeling the difference is within test variations, usually 2% or more likely the retest isn't an exact duplicate run of the 1st test. Unfortunately the only way to really get repeatable results with benching SSDs is ensuring that it's always writing to 'new areas'. Which is why some people go nuts on here over secure erase. Too bad they don't quite understand in the end that in real life after real usage it ends up resulting in less than a second of a difference. If you end up missing pressing 'ENTER' or mistyping a URL, there goes all the time you hoped to save doing a secure erase.

 

[Fedaykin311]

Your CPU is many orders of magnitude faster than any SSD on the Market (see difference between horse draw carriage and a scramjet). In a typical desktop setup, CPUs spend the majority of their cycles waiting for data to process -- even when all the data is cached in RAM. This is why SSDs are such a great upgrade, they upgrade that horse drawn carriage to at least a modern econocar -- reducing the #1 bottleneck in a desktop =)

However, the V3 actually does exceed the limits of most modern SATA controllers and perhaps even the latest and greatest SATA 3. So, the gains they are seeing, assuming they are real, is from the overclock also boosting the SATA controller.

 

[Emulex]

Yes C1E and other slow-down methods will cause a lower score under low speed situations. If you have your cpu at 100% all the time - you will have a marginal increase - probably not detectable but under the scope of a few specific benchmarks.


remember how you would overclock back in the days and it would bump the ide/sata speed up until it crashed the drive controller. One might ask can we overclock SSD? heh most definitely

 

[ksec]

It was a non English website, http://www.hkepc.com/6219

With Intel Core i7 2600k and P67, Vertex 3 can go up to 450MB/s.

If it is overclocked to 5Ghz, Write speed goes up to 516MB/s

 

[sub.mesa]

This should not affect throughput, but rather affect random IOps, especially random read IOps is affected by CPU performance.

Note that on Windows platforms, the I/O backend is single threaded, meaning it makes use of one core only. That means on quadcores with hyperthreading, you can be 100% CPU-capped with a cpu usage of just 12.5%. Keep this in mind and do more thorough benchmarks.

Note that disabling C1E may not be wise, since the benchmark does not combine the I/O with heavy CPU usage, like is expected in a real situation. So you may not really gain anything other than more reliable benchmark results by disabling C1E and other important power saving techniques.

 

[Emulex]

you sure windows server is single threaded? the more lun's i attach the more i/o worker sqlos dispatches you can even tweak the settings in soft-numa of sqlos. the x58 (mac pro) has dual nehalems which means each cpu has its own memory and bus.

so you assign one nic to each cpu and each cpu is numa so you want to keep all i/o attached to each numa cpu for the fastest i/o.

if you had one nic and two sockets you'd have to use QPI to do all i/o.

That doesn't seem logical.

 

[sub.mesa]

Well at least a AS SSD or CrystalDiskMark test on Windows would put 100% load on one core, and be CPU bottlenecked. This can be clearly seen on SSDs when doing multiqueue random reads (QD32/64); which would throw alot of IOps at the CPU where a single core starts being saturated. The IOps doesn't scale further, which it could have if more CPU cores were used by the I/O backend.

For mixed I/O or sequential I/O, the CPU usage is much lower and probably wouldn't be a real bottleneck anytime soon. Besides; which desktop application puts 200MB/s of random I/O sustained for more than a second? Doubtful, but on server systems this is much more common. SSDs used as L2ARC on ZFS for example.

 

[Emulex]

did you run As and CDM as the same time?

 

[Compddd]

My system has a Intel 160GB G2 SSD and becomes sluggish after awhile if I use C1E. After a fresh reboot, things like MS Word/Excel/Outlook will start up in under a second, but after a few days of use they will take 2-3 seconds to start up and the system seems in general sluggish.

Disabling C1E fixed this for me. Has anyone else run into this sort of problem?