I got to wondering recently if the Vista SuperFetch technologies were really working with me or against me when trying to get my UMPC to perform better.

Undoubtedly the drive in the ASUS R2H isn't the fastest performing thing around even when it is bursting a lot of data and what my experience of the machine being desperately slow when I closed applications made me wonder if SuperFetch was really working against me.

SuperFetch is supposed to prefetch regularly used pieces of code, beit system code or applications and load them into RAM as RAM is available... ergo unload something and the prefetcher loads something else that may speed up the user response if that application is called upon. (as I understand it).

Given that I increased my R2H from 768Mb to 1280Mb then just maybe SuperFetch is bringing a LOT of stuff into memory each time I close things.

Anyway I disabled SuperFetch entirely by changing..

HKLM\System\CurrentControlSet\Session Manager\Memory Management\PrefetchParameters\EnableSuperfetch

to 0 (from the default of 3).

The only other tuning that I have really is Memory Management\DisablePagingExecutive = 1 to keep all kernel and driver code in memory anyway.

I have found that with my config that performance is generally faster so I am curious if anyone else can verify this with R2Hs running 768Mb instead of 1280Mb and possibly other UMPCs.

Specifically I have found that operations like control panel navigation are much faster - this kind of thing loads and unloads a lot of CPLs (DLLs actually with a standard entry point) and I think that particular option keeps SuperFetch quite busy.

Other info : I did have a ReadyBoost compatible SD card (SanDisk Ultra II 1Gb) running and I am presuming that ReadyBoost is now also disabled since I think only SuperFetch uses ReadyBoost and with SuperFetch disabled the machine goes back to using the XP style prefetcher.

-Andy

How do you know that's faster?  is not a psychological thing?

Fewer applications not responding (actually slow to respond), less general disk activity, onscreen keyboard sending the correct chars (rather than misqueuing the same chars on repeat) and general subjective view.

It is hard to make an impirical observation with perf stats since we are measuring apps opening and closing etc.

I could measure disk IO and CPU but we know that they will be higher with SuperFetch enabled.

-Andy

Exactly, why you don't try to run sort kind of Benchmark tool. And then compare the results before the trick and after.  If you look around the web for this hack you will find a lot of contradictory information between people claiming that it works and people saying that it does not.

Please take the time to read what I have actually said ctitanic. What you are replying with is dismissive and all I am asking for is comparative views.

I have also already explained why a benchmark is of little value here.

-Andy

iamasmith:

Please take the time to read what I have actually said ctitanic. What you are replying with is dismissive and all I am asking for is comparative views.

I have also already explained why a benchmark is of little value here.

-Andy

I read it. But I insist, there are benchmark to measure the speed opening files.

OK, the speed of opening files actually isn't quite what I'm hoping to address here. Perhaps some background will help.

The UMPC device is somewhat resource constrained in terms of processor specification and possibly in terms of disk/controller/southbridge bandwidth. The main observation when the machine is running slowly is that there is often a lot of disk activity, particularly when closing and applications that does not even have data to write at that point in time.

Now I know all too well that small devices can be severely constrained in terms of disk controller bandwidth and have even worked on system code on devices that share northbridge and southbridge for the same function (memory on the same bus as devices being controlled). In these devices there is heavy contention for the bus and although that may not be the case with the UMPC there are some basics that still may be applicable to gaining a more workable experience.

The basics that I am talking about are that you don't max out certain resources in the normal operating mode.

The area that I am focussing on here is reduction of the IO load against the disk since this could possibly effect..

* Timely transfer of data for other operations in progress.

* CPU utilisation.

Even with DMA and Bus Mastering the CPU is still involved in many cases in manipulation of buffers sent that 'appear' in memory and orchestrating transfers so the low performance CPU is a potential real bottleneck here.

The specs of the disk controller and the bus bandwidth of the R2H? Not sure, haven't gone there yet. I'm at the early stages of speculation and often tuning focussed areas can yield faster results than a full understanding of all components.

Now given what I have said the only metric that would yield any form of evidence would be the Disk Queue Length and particularly that the Disk Queue Length sat at a particularly high level in general operation.

We should expect the disk queue length to be higher on a low spec macine anyway with SmartFetch enabled since it is working all the time. The question is really is this an efficient way of working since although loading may be faster once things get cached back, is the interval in between whilst SmartFetch is constantly hitting our constrained resources at the general detriment to the experience.

I will run some tests on queue length with SmartFetch enabled. I anticipate they will be higher than without it enabled but in terms of what I am attempting to convey I feel it proves little since the efficiency of the cache read has to be offset against the level of disk queue.

The only thing that the queue length may suggest is that if it constantly exceeds an unacceptable level then it is a true bottleneck.

Anyway, I will give this a try when I have a little time.

-Andy

Thanks Andy! It´s not that I do not believe you, it´s a matter of having something to see if that realy works or not. Thanks again for your time on this matter.

I had a 2GB SD card as readyboost with 1024mb of Ram, but it seemed that the readyboost just slowed my system down. Up until 3 minutes after restart a service was taking 70% CPU, this dissapeared when I disabled ReadyBoost.

I've since then reverted back to XP but when my Upgrade arrives I'll try again..

Now that's another interesting point.

When I get to do the tests I will test with and without ReadyBoost.

-Andy

OK, here are the stats back from the rudimentary test.

Note that loading and unloading the control panel didn't actually seem to make much difference. Unloading a large(ish) application - in this case Outlook 2007 - did make have some effect.

With SmartFetch disabled I saw mostly a Disk Queue Length of 1 but with peaks as high as 3.

With SmartFetch enabled alone I saw a significant amount of activity and there was a prolonged period of activity until the system stabilised and disk queue went to 0. During this time the Ave Disk Queue Length hit 5.

Unloading Outlook made the Disk Queue hit 5 for about 11 seconds solid.

Now the interesting thing. Add the ReadyBoost device to the SmartFetch enabled config.

The average never went below 1.25 but an unload of Outlook made the queue peak quite heavily hitting 6 for a some of that period for a period of about 45 seconds! - this is the kind of performance issue that made me start looking for a cause.

CPU utilisation was also pretty high on the SmartFetch enabled configs  97-100% not uncommon for prolonged periods.

Anybody else feel like trying this out and seeing if this may be linked to the increased amount of memory or a general thing on these relatively low specced machines?

I should mention my config at this stage...

Office 2007 Home and Student - OneNote is running as a tray app.

Outlook 2007 (man it's expensive to add Outlook for an end user Consumer - my business pays about 3 x this price for Vista and Office Pro Plus with Exchange and TS licenses per user :(()

Kaspersky AV very latest version - running through all tests.

Nothing else.. I like to be lean and mean :)

-Andy

Good work Andy! I´ll test it tonight!

Quick question, all your tests were with Memory Management\DisablePagingExecutive = 1

Yes, all tests were with DisablePagingExecutive = 1 since I figure that I have the luxury of extra memory over other users.

Another thing I forgot to mention. I'm using Vista Business so I disabled the system recovery to stop the Volume Shadow Copy potentially performing additional work.

-Andy

I just tested it in my Q1 and I really feel it faster. I know that your tests are a lot more efficient to measure any changes in how the machine performs but to could resist andrun a Benchmark tool:




Before changing the Registry



there is not a big difference but the fact that there is a positive difference already means something.