Tweaks For benching disscussion Thread

I created this thread to talk about tweaks and what they can do to make a difference when trying to squeeze the most out of your runs.

I am hoping that this will help those including myself trying to
achieve the most out of a benching session. I know Snod and Hyena will be able to help as hyena has written articles on this subject and Snod can contribute with all the research and knowledge that he has this subject.

I want to show an easy way that I cut Super Pi 32M run down with some simple OS tweaks and a very broke down version of wazza that I have found to work. While I do know about eram and such I am just now beginning to get into tweaks and trying to understand them so bear with me.



I am sure that others here will be able to add to and explain better what I am about to show.

I used my laptop for this example to show what kind of time you can shave off w/o ocing since this is about tweaks and not Ocing.

I ran a 32M to show off a starting point which I will start from.
(Each of these runs were done 5 times and the best was taken from each set)

32M No Tweaks-1625.172s



I then cut down on the processes and background services by running windows in diagnostic mode.
I also changed the visual effects to best performance and switch ed the memory usage from programs to system cache. I then ran 5 runs of 32M and this is what I came up with.


32M Basic OS Tweak-1563.094s



That shaved 62 seconds off the benchmark. Not too bad,eh?

Now there are tweaks referred to Copy Wazza.

Here is a quote to help explain it.



Here is a video to better explain it and how to set up eram.
------------------------------------------------------------------->Link to Video

Now you find that there are ridiculous amount of variations to this and everybody has there own little way of doing it. I myself did not have two HDD's on my lappy to do it as described so I tried something a little different to see if it would work.


32M Basic OS Tweak & my tweak-1539.219s



Took off another 23 seconds.

As you can see just by killing some services and learning the wazza trick you can shave off some serious time on your 32M run. In my case I was able to take off a total of almost 86 seconds with any OCing at all.

Here is a quick shot of what my time was after tightening up my ram....


32M Final Run -1495.438s



A total of almost 130 seconds.

I am going to be spending time learning and doing research so I will be sure to share anything I find out. I will show you guys the "half ass wazza" tweak I performed in order to better understand what I did and to give any feedback on it.

Well there's always the well known maxmem tweak and there's something people keep calling LSC=1 but I don't know what that is? People keep these things much too quiet, haha..

I must let it be known I've never used any of these though, all I usually do is move everything off the desktop into a folder somewhere else, as apparently things on desktop are cached and cost you speed, remove the background image, and disable as many services as I can without getting the wonderful 60 second timer..

IIRC, LSC=1 is where the tweak is done in the registry for Large System Cache concerning memory. The equivalent would be setting it in the system options, if what I am talking about is what you are referring to.

The maxem thing....

I understand the principles behind it but as soon as I do a little more research to back up what I am currently doing yields the same results w/o all the extra steps that everybody post for copy wazza. It just so much dam time testing

I spent at least 6~7 hours today rerunning 32M w/ different settings.

So - here's the SuperPI excerpt from my article.

“First tweak: We enthusiasts call it the "MAXMEM" tweak. It basically lowers the amount of physical memory that Windows is allowed to use for system processes. This tweak often slows your system down but ultimately yields better results with Super PI benching.

Apply this tweak only if you're aiming for the top and want the absolute best. Here's how to enable it. Use a text editor such as notepad to open C:boot.ini manually or go to Start -> Control Panel -> System -> Advanced -> Settings -> select your operating system (that's the default one, usually) -> Edit. Now go to that line that ends with the '/fastdetect' command. Append to the end of that line the following command: "/maxmem=104" (without quotations marks). It should look similar to the example below:

multi(0)disk(0)rdisk(0)partition(1)WINDOWS="Windows Server 2003"

/fastdetect /maxmem=104

Save the changes, close the file, and then reboot. Experiment with this value. Apparently 104MB leads to the best results for a Super PI 1MB run but 600-650MB is the best option for a 32MB run. So when you plan to run Super PI on 32MB, don't forget to set an amount around 600MB, let's say 625MB. When you're not benching at all, delete that key to use your maximum physical memory.

Second tweak: We call it "copy-waza." Actually this tweak is originally from Team Japan, the mighty gods of over clocking; kudos to them. This tweak yields the best performance on 32MB Super PI runs. On 1MB the difference is usually unnoticeable. The technique works like this: you copy a very large file from one partition to another one. Then, exactly as soon as the copying process is done, you hit the "OK" button to start benching.

Why does this work? OPB (Onepagebook), one of the best overclockers in the world, performed a few tests regarding this tweak. Hundreds of others also tried and tested it. They found out that the process of copying very large files lowers the amount of memory allocated for the "explorer.exe" process from around 18-20MB to around 8-10MB.

It's both funny and weird. You're skeptical? Just give it a try yourself! Again, the results may not be astonishing at first glance but, hey, every little improvement helps. Why? Because applying all of these at the same time in a perfect correlation often gives you that must needed one-second or half-second difference.

One last note on this tweak: The "very large file" I've been talking about needs to be at least 2GB but 3-4GB works the best. I've personally been moving around (cut-paste also works, because it's copy-paste-delete) a DVD image file (~4.7GB).”

Sure, the things I mentioned might have been already outdated in terms of there is a possibility that lately there are other newer tweaks and tricks that yield better results. In terms of brute-forcing tweaking what once worked, chances are they will always work, just the gains can be much lower.

LSC stands for Large System Cache. It is disabled by default in Windows XP. It is enabled by default in Windows 2003 Server.

LSC = 0 means that it is disabled and then this represents the following: Maximize Throughput for Network Applications, which uses cache memory only for I/O functions.

LSC = 1 means that is is enabled and then this represents the following: Maximize Throughput for File Sharing, which uses program memory for I/O functions when the cache is full.

Technically, what large system cache does is enabling the operating system to use up to 80%+ of system cache during normal conditions. This is from where comes the boost while benching on extensive calculation-based applications, or this is ideal in case of web servers. In everyday applications it should be enabled since foreground applications are ought to have priority to use RAM instead of system cache. There's no need to involve system cache until it isn't necessary and/or explicitly required to do so.



In case of Server Operating System (but perhaps even XP - but I haven't got a clue) you can tweak this further using the following registry key: "HKEY_LOCAL_MACHINE\SYSTEM\CurrentControlSet\Servic es\lanmanserver\parameters". You can set 1 (Minimize Memory Used), 2 (Balance), 3 (Maximize Throughput for File Sharing and Maximize Throughput for Network Applications).

Also, I'd like to point out something regarding Copy Waza. Brian's instructions are using Microsoft's "fsutil" method/command. Using its "file createnew" parameters indeed a new file is going to be created with the specified file size. However, what I personally dislike in this is the following: "Creates a file of the specified name and size, whose content consists of zeroes." - that's the definition of the "createnew" parameter as it's stated at Microsoft's Fsutil: file documentation (link: here).

I have tried it myself and indeed the file was filled with bunch of NULLs (white spaces). That's not good on the long-term. I don't know whether the de facto standard Windows copy/paste function is optimized to check the file at first or just it plainly copies the file regardless the file, but in some cases it might be "smart" enough to realize that the file has nothing but white spaces and then the copying feature is done faster (more efficiently) using different algorithm(s) than just vanilla linear old-fashioned copying (copy bit X from A to B in place of Y) and that's NOT what we need.

We need to put heavy load on explorer - that's why Copy Waza works the best. That's why I have recommended in my article copying files around 4GB-5GB and this works the best with DVD image files (4.7GB) or who knows, anything that's one long file and its lenght is that much. Or, if that's what you want - just ask either me or Yeti and we can write a small console-based application that creates a new file and fills it up using random bits so that no kinds of optimizing algorithm can be used to effectively copy the file other than just linear copying.

This is crucial in my eyes. Sure, Microsoft's copying function might not have been optimized at all and if this is the case then I'm wrong - but before doing extensive tests or having proofs on whether or not the copy is optimized or not, we cannot know with absolute certainty and it's better to avoid possible consequences/drawbacks and applying a technique "wrong" way when there could be achieved bigger edge/gain.

As well as I'd suggest closing the process "explorer.exe" - re-starting it from create new task "explorer.exe" (if it does not re-start automatically), and then starting the Copy Waza trick. This is also useful because over time the Windows Explorer might have allocated (dynamically) lots of unnecessary memory fields/bits. A clean re-start should free these up and you can do this by killing the process and re-opening again.

The next tweak that I think is definitely worth researching - and most importantly playing around with - is eRams or RAMdisks. I have used Ram Disks ages ago and they do give boosts. The utility/application that I have used was developed by a guy called Igor Pavlov, a Russian software developer, as far as I know the project has been discontinued and it is not officially supported anymore. Needless to say, I have hosted in case you want to give it a try - it has extensive and straightforward documentation (included in the archive in PDF format) - and managing this virtual RAM disk is anything but black art. Download it from HERE and give it a try.

I would personally recommend placing SuperPI (as per the files) on this virtual drive. I have further ideas on optimizations such as creating a work-around that emulates system disk cache - which in fact is nothing but a virtual RAM disk - and then limiting the amount of available RAM (using /maxmem) and the rest would stand for RAM disks (for system cache and SuperPI). All in all, you would force Windows using disk cache more often than necessary (LSC=1) which would also give enormous speeds thanks to RAMs bandwidth and tight latencies (and memories are ranking on way better spot in the cache hierarchy compared to HDDs). Ultimately, you could also play it around and if you have 2GB of RAM then another further optimization would be setting up the paging file/swap space on a RAMdisk too. Think about optimizations. The key here is to physically limit your available RAM with /maxmem and then you should allocate and use up the rest with RAM disks.

Hope it helps - I've got a few more ideas but most importantly we'd need a lot of time to test because we cannot know which would give significant gains or not. You know, good old-fashioned trial-and-error.

Cheers, Brian!

Hey MAD you should of put some b-4 and after pics like Brian so we could see a diff