Is prefetch bad for SSD's?

Here's some testing I did a while back with 2xSSD RAID0 and a 16MiB executable.

The way the memory manager works is to load the first 4k-32k of the start point of the executable. The executable starts running then when it jumps or tries to access memory outside of that 4k-32k a page fault occurs and the 4k block for that section has to be fetched from the disk into memory to continue operation. This keeps happening until all 4k pages on their first access have been loaded into memory.

Running with prefetch disabled.

Access to disk can be seen starting off at 64 sectors at a time (32k). As memory access by the program can be somewhat random it becomes fragmented which means it can drop down to as low as 4k. This results in first time access of somewhere between 32k sequential and 4k random speeds. Launching is fairly quick, just under 10ms but execution time takes over 202ms. About 200ms of this is disk access time.

Running with prefetch.

The executable is loaded first in blocks of up to 4096 sectors (2MB) before execution takes place so understanably the launch time is longer. Note how much quicker disk access time is, 30ms vs 200ms. The executable itself is free to run without hard page faults and finishes in 2ms vs 202ms

First time running with prefetch is over 6 times faster than without prefetch.

 

Thank you for providing the 'numbers' behind my 'feeling' that a system with PreFetch enabled is faster.

Every time I mention this to anyone; all I hear is how NOBODY can feel less than xxx milliseconds differences.

I can and do and the proof is above.


Dufus, if you're able to report this for us; imagine what MS is doing under the hood to improve each Windows O/S as it goes forward.

Another example of not needing to mess with SSD 'optimizations', or otherwise - they usually hurt, not help the system in question (overall).


Windows 8.1 has so many of these built in improvements that when I'm forced to run a Windows XP system (to simply load my delivered images to a client) I can almost see the old XP's gears turning as it figures out what it has to do next to simply copy and display the images (and this is still on a Core 2 Duo with 4GB RAM and XP Professional SP3).

I think the moral of the story is: don't second guess the software engineers that have created the O/S (sometimes, they know what they're doing).

As for how bad it is for SSD'? So insignificantly small that it is ridiculous to even give it a passing thought. Far, far, FAR worse is running an SSD without OP'ing - that is what actually kills nand cycles as WA factors climb into the double digits and performance over time plummets to below HDD speeds.

 

Hi,
I can not find Prefetch service in my 8.1( Samsung 470 ) and SuperFetch is disabled by default. I haven't done any tweaking and been using as it is. My question is whether Win 8.1 disables SuperFetch when it detects a SSD?

 

No, it does not. However, if you installed the Samsung Magician software, depending on the settings, it will automatically disable it. I'm not sure what the default is, I know that the Intel SSD toolbox will nag you about tweaks which include disabling superfetch when you run it for the first time, but it is left enabled by default.

 

Which begs the question... which one? MS or SSD manufacturers know what they're doing?

 

Thank you guys, I am going to enable SuperFetch and see, actually I am not bothered about the longevity of the drive as I might buy another drive long before mine dies

 

My money is on the SSD manufacturers ... Windows 7 & SSD: defragmentation, SuperFetch, prefetch (just in case anyone want to disable both of these in Win 7)

 

Neither.

You do realize that that page is from November 2012, right? Two new O/S's have been releases since then. (Win8 and Win8.1 ).



The one who is right is the one who has tested many different setups, first one way, then the other for up to a month in each configuration - and has seen firsthand that disabling SuperFetch, Pre-Fetch, Indexing and other core Win O/S technologies is simply getting a slower system in the end.

So who's right? I'm right!

 

My own personal guess: both are aiming for different things. Disabling superfetch/prefetch will save you some writes and in that sense it will increase the longevity of the drive somewhat. On the other hand superfetch/prefetch improve certains aspects of performance. I would say it all depends on what exactly you value and if you notice a difference in launch time for your files/programs with superfetch enabled/disabled.

Now, disabling the default Windows defrag, I completely get. MS's article that JOSEA links states that superfetch/prefetch are there mainly to improve performance on a HDD and that on a SSD, since it's fast enough (entirely subjective metric there btw), it serves no purpose to leave it on and that it should be disabled to avoid extra writes. That in part brings me back to the first paragraph about that's fast enough and what isn't and that it will vary according to different people.

In the end, it all boils down to what everyone should do imo: learn the basics of the subject and make your own informed decision according to your needs/wants.

 

True...but not everyone uses W8 or 8.1

The real question is whether the linked article still holds true for W7.

 

Windows prefetch can be a bit hit and miss sometimes and could be improved IMHO.

Yes they have brought in new API's with W8 with up to 8 levels of prefetch. Server W8 even gets something extra.

CLR - An Overview of Performance Improvements in .NET 4.5

Note the "during application startup" not to be confused with the old way superfetch worked by pre-empting applications and loading them into memory cache before they were required even if they might never be run.

IIRC MS decided SSD access was "fast enough" to not warrant use of prefetch with SSD's which doesn't mean access couldn't be faster. This thread is mostly just for information purposes only so I hope it doesn't turn into one of those "shove it down your throat my way or the highway threads". What you decide is right for yourself IMO is fine.

 

This might be a dumb question, but for someone who isn't concerned with SSD longevity and simply wants the best performance possible, would you recommend enabling both Prefetch and Superfetch?

 

http://forum.notebookreview.com/sol...age/744191-prefetch-bad-ssds.html#post9552729

 

Gotcha, will enable both and see...

 

Just note that it is not something that you will notice in an A/B back to back quick test. Nor with normal benchmark tools (that I have seen), the exception being this very thread started by Dufus.

This is what the experience will be when enabled one way or the other over a week/month (depending how long you take to start a repeat of your workflows and allowing for normal/regular idle time in between using the system and turning it off) of your normal use.

Then switching the setup and seeing how your next week/month is.

Each time I have done this (with Arrandale and Windows Vista and Win7, SNB and Win7 & Win8 and IB & Haswell with Win8 & Win8.1), the outcome was the same; having these disabled made for a comparatively sluggish system each time.

A couple of more caveats; the proper (and best) drivers need to be installed for each device on the platform - including the latest Intel RST SATA AHCI or RAID drivers that allow these subtleties to be apparent. Do not install the IRST driver with the exe file; instead; install via Device Manager with 'update driver' and point to the appropriate x86 or x64 'F6 floppy' you downloaded, depending on your installed O/S version.

 

rhutor, from what I have seen while W8 reports a drive as an SSD it will not use prefetch with that drive even if it is enabled and a prefetch file exists for the application. Due to Windows not seemingly using the identify data of a RAID array it can be reported as a hard drive even though it uses SSD's and then prefetching from SSD can work. This is how the example posted earlier was able work and take advantage of the higher sequential reads of a RAID0. I don't know if trim is still sent in this state but defrag can be disabled and the array trim optimized via the -l defrag option. This is all far from ideal for everyday usage though but hopefully MS might bring some options later on.

 

Dufus, that is interesting.

Does that mean that SuperFetch works the same way then?

 

How exactly superfetch works these days I do not know. I personally was not so keen on it's initial use of pre-loading commonly used programs whether they were going to be used or not, but would like to have the option to prefetch when the application is launched.

 

Hmm, so it appears that Samsung's RAPID mode is akin to Superfetch. Is there any point in using Superfetch in that case? (actually, are the two even compatible with each other??)

 

No, RAPID is simply a third party RAM cache. I highly do not recommend it.

 

Isn't Superfetch also a RAM cache? I'm guessing you're saying no to RAPID and Superfetch at the same time?

In any case, I had RAPID enabled for a while previously and can honestly say I couldn't feel any difference apart from getting better benchmark scores. Pretty much the only reason I left it on was for benchmarking.

 

SuperFetch is technically a (smart) RAM cache. But the difference is that the O/S supports it like a first class citizen at a core level. Highly recommended whether you're running HDD or SSD.

While RAPID is just like FancyCache, etc. with the probability of crashing and thrashing your O/S: HIGH in my experience.

The best third party RAM cache I've used was eBoostr (reliable and made a difference) - but as soon as I discovered the consistent performance of SSD's with ~30% to ~70% capacity left as 'unallocated', even eBoostr was relegated to the past for all my setups.

 

Turned on Superfetch and preftech, boot time increased by 5 seconds (from 15 to 20). Yes I know it's "only" 5 seconds but that's a 33% increase! Off to bad start...

 

Alright it's been 2 week since I enabled Superfetch and Prefetch, and I can honestly say I have not noticed any performance improvement whatsoever. Applications do not seem any snappier than they already do -- if anything upgrading from 1600 to 1866 ram made a bigger difference. I do have the latest IRST drivers (12.9) as well.

The only noticeable change was that boot time increased by about 5 seconds. Nothing to write home about, but still frustrating nonetheless. I have since disabled Superfetch and Prefetch after giving them a fair shake. Oh well.

 

How much RAM do you have? Is system restore enabled? Is error reporting enabled? Is hibernation enabled?

How long is a typical session with the computer? How long is the computer left on idle, normally?

Is the pagefile disabled?

The way you have setup (including hardware) and use the system will make a difference.

 

16 GB ram, system restore disabled, error reporting on, hibernation disabled, pagefile set to 2048 MB.

Typical session really depends on the definition, but I turn it on after I get home from work, say 6pm, and shut it off around midnight. Computer is normally not left on idle for long -- I'm either browsing, gaming, or watching movies.

In any case, at least for my usage pattern, I saw no benefits in leaving Superfetch and Prefetch on, so I disabled them.

 

Thanks for replying. I'm curious if the pagefile is slowing you down? Are any programs requiring the pagefile at all?

 

I keep the pagefile because I run some legacy DOS games with DOSBox, and also didn't feel entirely comfortable with disabling it completely.

 

The only way to get comfortable: close your eyes, take a deep breath and jump! (Turn off the pagefile).

Be in test mode for a couple of days and then decide if it works for your workflows or not.

I too was wary - but especially with Win8.1, disabling the pagefile has 'always' been beneficial for my setups (and 16GB RAM or more).

 

rapid mode just like a ramdisk you cant use, but way way less useful lol.

i'll try enable superfetch and see if there are any performance increase. as a gamer at competitive level I am able to tell a few ms delay difference =) also running multiple laptops on win7, 8, 8.1 and server 2012 R1 and R2. 8 and 8.1 are technically the same as server 2012 R1 and R2 though

 

RPAID is useless except for getting better benchmark scores. Just like Super/Prefetch it also slowed boot time by a few seconds without an iota of noticeable improvement in daily usage. I have Win7 so it might not be as optimized as Win8.

 

it is recognized by many reviewers win7 will give better performance to storage than win8. server 2008 r2 places first, then 2012, then 7

 

I'd like to see those links to those reviews.

I find total time to complete a real world project (boot to shutdown) is faster with Win8x64 Pro than any other O/S - including the storage responsiveness and absolute performance side of things.

I am network based though - so that might be skewing the results I am seeing. Still even if Win7 is faster, it may only be by a few milliseconds (I guess) in a real world workflow.


See:
Confessions Of A 10 GbE Network Newbie - Part 1: Basics - SmallNetBuilder


The above link (I hope) shows why Windows 8/8.1 x64 is superior to any previous O/S where the storage subsystem over Ethernet is a concern (NAS/SAN etc.).


Still, I would be interested to see any claims to the contrary.

(Sorry, can't seem to edit my post above).


As if to underscore how limited Win7 is for networked storage when speed is of paramount concern; this article makes me smile.


See:
400GbE Networking Standard Ready For IEEE



50GB/s connections between computers will make even the fastest m.2 based PCIe based SSD's look weak in comparison.

And will further differentiate Win7 vs. Win8.1 and all future versions going forward.

 

unfortunately reviews has been taken down cause website no longer exist. Rwlabs basically move everything to TT but they don't have all the articles anymore. but tests were done long time ago as early as maybe late 2012 I think, and yes windows 7 is only a bit faster than 8 probably only noticeable in benchmark. server compare to windows 7 on the other hand shows even more performance increase especially on 4k write in benchmark. if windows 7 you'll get around 400MB/s in 4k write in raid 0 of two SSDs, in server 2008R2 you'll getting maybe 600MB/s which is a pretty big increase, 20-25%.

heres some old forums about the links it use to exist, was done by Jon Coulter. same guy is now at TweakTown doing raid reviews. he also tested a bunch of raid driver and OROM combination for best performance which leads to 11.2 driver being the best, for raid.

SSD Reviews Thread | Page 7 | TechPowerUp Forums
look for SuperSSpeed S301 128GB SLC Win.7 VS Win.8 SSD Review

 

Wait, 600 is only 20-25% larger than 400?


Sent from my iPhone using Tapatalk

 

I spent a good ten minutes looking for the article - but no dice, they're gone from the web.

400MB/s 4K R/W scores? Even from two drives in RAID0 that seems almost impossible? Not 'real world' in any case.

Even if the articles were available; that would have been early (beta) Win8 performance numbers not to mention IRST v9 or 10 at that time.

I'm talking performance from last week; Windows 8.1 with all updates, IRST 12.9 WHQL or beta v13 is untouchable with regards to SSD performance for real world workloads. Especially with network storage - any other O/S seems weak by comparison in these metrics.

I used to be a little jealous of how fluid OS/x was - not anymore. The latest OS/x is Vista for Mac (TM ). Sluggish, cartoonish and not unlike IOS 7.x now that I think of it. In other words: both (OS/x and IOS 7.x) are useless when maximum responsiveness, absolute performance and longest battery life is needed.


Want to laugh? I suggested a Haswell based i7, 16GB RAM 480GB SSD Windows 8.1 Pro based system for a clients secretary's replacement computer. They're not sure if it will be powerful enough (they're basing this on price: they were quoted ~$4K for a new system, including monitor, keyboard, mouse, etc.).

They're currently running an XP Pro computer (was a steal at $7K when dinosaurs roamed the earth) with a single core cpu and SATA1 (yes; 150Mb/s max connection) dual HDD's (which can be heard outside the office and around the hall). This setup has 1GB RAM and can open a network database in less than 5-10 minutes (depending on how long the computer is left idle for after it turns on). The size of the database? A whopping 2MB (yeah; an mp3 file is bigger).

The quote for their new 'underpowered' computer is coming in at just under $1K. Do you think the secretary will like it?


(Probably not; less time to keep herself looking like a million bucks on her boss' dime - polishing her nails, checking her lips, etc. etc.). lol...



Back on topic:


What two SSD's in RAID0 do 600MB/s 4K R/W? Are these PCIe based?

Also, yeah; 600 is 50% larger/faster than 400...


As for those old reviews; doing a 'first look' at things is great (we're all curious, of course). But the underlying tech that a new O/S has needs time to fully blossom. Windows 8.1 x64 Pro has hit it's stride, imo.

I'd like to see a good review today of all the major O/S's with regards to SSD performance. Not benchmarks; just someone with 3 or 4 identical (hardware) configured systems (except for the O/S) and the subjective 'feel' of using the systems for day to day tasks.

Not that such a review would make me switch to another O/S (I know what works for me already).

But it may push the O/S manufacturer I prefer to use even harder to keep improving what we'll be offered with Win8.1.1 or Win9 very soon.


When Intel (circa 2009) was saying that the new cpu's and platforms coming out would need SSD's to maximize their performance, I was skeptical.

When I see what an Atom Z3740 can do with even a so/so eMMC SSD and Windows 8.1 x86; I can easily see what they were talking about.

See:
Asus Transformer Book T100TA, 64GB w/ Keyboard Dock at Memory Express


When you consider that not even that level of performance was available almost a decade ago for several thousand dollars more (not to mention being just the 50x bigger beige box without monitor, speakers, keyboard, etc.) and the T100TA lasts 8Hrs (or for my use: 3-5 days checking email), you can see how far we've come.

And at a price that would make the circa 2003 manufacturers cringe, I'm sure (at ~400 on sale).

 

4k read from my SLC S301 goes as high as 58MB/s for incompressible file dependent on my raid block size. this is one of the biggest advantage to sandforce controller paired up with SLC. I have my thing set to 8k stripe to reach 58, could go for 60~62 if set to 4k but loses out on sequential performance.

4k write is the one thats affected by OS difference the most. win7 at 400, server 2008 R2 at 600.

when Jon does review next time on Tweaktown just leave comment there asking about differences in OS, he'll tell you. server is definitely faster than windows 8/8.1, cause server is MS's best which make sense. 8.1 may or may not be better than 7 that I havent test but 8 is a tiny bit slower than 7. just like server 2012 is a bit slower than server 2008 R2 since 2012R1 is based off 8 compared to 2008R2 based off 7.


can't wait for 256gb SLC if they ever do release one.. with sandforce 3000 controllers

 

unityole, I still want to see 600MB/s writes at 4K R W.

Any links? Anything I searched for showed the same old same old; ~50MB/s.

Are you talking Mbps by any chance?

 

600MB/s 4k? Wowza. I doubt it. It's phenomenal to see 60MB/s

 

lol are you guys having problem reading what I said?

I said.. 58MB/s 4k random Read, 600MB/s 4k random writes, in server 2008 R2. the 50MB/s you see is probably with 32-64kb stripe size in raid 0, tested by Jon, I used 8k stripe size, I can reach 62 with either 3 drives in raid 0 at 8k stripe, or two drives in raid 0 with 4k stripe but I'd lose out some sequential speed.

sequential isn't all that much on chipset nowadays anyway, haswell H87 is only able to do 1600MB/s read I think, HM77 in my alienware only able to do 1450MB/s at Queue depth of 1 read. writes are even slower, for sequential.

 

Lol... I think you're having troubles writing what you want to say?

I'm still seeing those as unattainable until proved otherwise.

 

or that, well if you had read Jon's review then you'd know. he did a comparison of windows 7 and 8, and then comparison in rwlab's forum for windows 7 and 2008 R8. the 50MB/s 4k random read is only for S301 SLC two in raid0 but thats stripe 64k that he did.

heres one from CDM run, windows server 2012

 

I have gone down the 'tweak the stripe size route' and more. In the end, it broke more programs than it helped.

Benchmarks are not what I'm about, but thanks for showing that one (I guess now, it's real).

What isn't real is using 12% of a secondary RAID0 drive array (meaning; not an O/S system drive, and one that has not been booted from, I'm assuming) with a tweaked stripe size and a SF based SSD drive tested with 100% compressible data.

In theory, sure; it flies. (You have shown the 'proof').

In the reality though; I would bet it was probably worse than a single SSD setup in at least one metric.


What tuning the stripe size taught me was that the overall performance of a system would tank even if the data was handled optimally (as indicated by a benchmark).

The other thing it bore into my thick skull is that today's O/S' use the hardware just as much as we do (programs + data). When you optimize the storage subsystem solely for the data and/or the intended workflow (think video, database, etc.); the snappiness and responsiveness of the system always suffers. Even with an SSD.


Dorothy, we're not in DOS land anymore.

 

the thing is even with incompressible data I get similar performance wit the graph I shown above, because it's paired up with SLC. this was the reason I am so hyped about any upcoming sandforce 3000 series controller paired with SLC flash from superSSpeed. I'd hope the next one I see will be 256gb not 128gb version, so getting a 512gb with 20% OP on a SLC would get amazing speed and I'm sure I'd feel the difference.

I am paranoid at this point right now I'd throw most of my stuff into ramdisk because it is way more snappy, majority of the stuff is browser profiles and cache files and let me tell you it makes huge difference lol, raided 0 SSDs can't even come close to ramdisk.

 

unityole,

I'd love to see this for myself (i.e.; my workloads), but right now I reserve my right to doubt!


Ramdisk?

Will make you crash faster. lol...

 

the same SLC S301 SSD review still exist on Tweaktown done by Chris there you can look at the incompressible files performance done with CDM. I try to avoid CDM and only really use it for 4k random speed test since it likes to pick the best results somehow (from my experience), also the sequential performance is at 4 queue depth so I avoid it.

how does ramdisk make me crash faster? given if all ram timing are stable etc.

 

When (not if) the computer will crash because of using the ramdisk - that very speed will make it happen faster.

(It was just a little joke, but the stableness of a system using a ramdisk is not dependent on ram timings: the software/workload is what will make it crash).

 

I'd assume computer crash from time to time due to ram, is it because that reason? I have left my computer running for months and never really had any problem though.