TweetFollow Us on Twitter

January 92 - Blueprint for Automatic Segmentation

Blueprint for Automatic Segmentation

Alan Bommer

Segmenting an application can be tedious and frustrating. Since MacApp eliminates many tedious and frustrating tasks for programmers, segmentation seems even more odious to users of MacApp. This article outlines two ways segmentation could be automated in MacApp.

The Goals of Segmentation

There are generally four (sometimes conflicting) objectives in segmenting a MacApp application:
  • Minimize the "temporary memory reserve." This reduces the amount of memory that is necessary to run the application. To achieve this goal, make sure the segments loaded at the time of peak temporary memory usage do not contain routines that are unnecessary at the peak time.
  • Minimize the time used for loading and unloading segments from disk. This improves program performance. Smaller segments load faster than larger segments, but loading two 5k segments takes longer than loading one 10k segment. To minimize the loading and unloading time, segments should be large, but should not contain routines that are unnecessary.
  • Minimize heap fragmentation. This speeds the performance of the Memory Manager and ensures that no memory is wasted in memory fragments too small to be useful. Larger segments minimize the number of handles in memory and hence minimize potential fragmentation.
  • Minimize the number of jump table entries. This improves program performance as intra-segment code-to-code references (no jump table involved) are faster than inter-segment code-to-code references that use the jump table. Keeping the jump table size less than 32k (4096 entries) can also improve the program's performance by eliminating the need for (the slower) "32-bit everything." Larger segments minimize the number of jump table entries, because they appear as intra-segment references instead.

The Statistical Analysis Approach

The statistical analysis scheme for automatic segmentation consists of three steps:
  1. Modify the source code so that every routine not in a user-specified (or MacApp-specified) segment is put in its own segment;
  2. Run the program a representative number of times and collect statistical information on the usage of segments;
  3. Analyze the statistical information and generate segment mappings (this step is by far the hardest).

Modifying the source code (step 1)

You must segment all routines in your source code in either the standard way-{$S segname}-or as {$S autoseg}. An MPW tool will then (by referencing Appname.MABuild) change all the {$S autoseg} directives to {$S autosegN}, where N is a unique number for each routine. For repeatability, the tool also renumbers any {$S autosegX} that it encounters.

Collecting the statistics (step 2)

The modified source code must be built with the options "-NoDebug -AutoSeg -ModelFar." MacApp will collect the necessary statistics by adding a new procedure and a few lines to UnloadAllSegments. Every time UnloadAllSegments is called, the new procedure will update a data file of a format similar to these quasi-Pascal records:
UsageRecord = RECORD
    flags: LONGINT; {is segment resident? etc}
    segmentSize: LONGINT;   {code size}
    usedWith: ARRAY[1..numSegs] OF LONGINT;
END;

DataFile = RECORD
numSegs: LONGINT;
usage: ARRAY[1..numSegs] OF UsageRecord;
END;

The "usage" and "usedWith" fields are defined so that "dataFile.usage[i].usedWith[j]" is the number of times that segment "i" and segment "j" are both loaded between calls to UnloadAllSegments.

This data file is very large. For 1000 segments (routines), the size is about 4M; for 4000 segments (routines) the size is about 60M. These sizes can be cut in half by taking advantage of the symmetry of the data file (dataFile.usage[i].usedWith[j] = dataFile.usage[j].used With[i]).

Analyzing the statistics (step 3)

The hardest part of the automatic segmentation scheme is analyzing the data. Empirical rules determine which segments were mapped together. Below are some rules in order of preference:
  • Limit segment sizes to 32k unless the "-modelFar" option will be used.
  • Segments that are always loaded together should be in the same segment.
  • Non-resident segments should not be mapped with resident segments.
  • Segments with the highest percentage of being loaded together should be mapped together before segments with a lower percentage.
  • Segments loaded more often should be mapped before those segments loaded less often.

The Total History Approach

The total history approach consists of three steps similar to the statistical analysis approach outlined above: (1) the first step is exactly the same, (2) step two is the same, except that the information stored on disk is the (almost) total time history of all segment loads, and (3) the third step is to analyze the history and create segment mappings to meet the goals explained above.

Collecting the data (step 2)

After the source code segmentation is modified with the MPW tool as explained in "Modifying the Source Code" above, the application must be built with the options "-NoDebug -AutoSeg -ModelFar." MacApp collects the necessary statistics by adding a new procedure (different than the one in the statistical analysis approach) and a few lines to UnloadAllSegments. Every time UnloadAllSegments is called, the new procedure updates a data file of a format similar to these quasi-Pascal Records:
SegmentNumber = INTEGER;

SampleRecord = RECORD
    numNonResSegsInSample: INTEGER;
    {system use of reserve (in bytes)}
    nonCodeRsrcUsage: LONGINT;
    {total use of reserve (in bytes)}
    totalCodeReserveUsage: LONGINT;
    segmentsLoaded:
        ARRAY[1..numNonResSegsInSample] OF SegmentNumber;
END;

DataFile = RECORD
numSegs: INTEGER;
numSamples: LONGINT;
sizeResidentCode: LONGINT;
peakCodeReserveUsage: LONGINT;
segmentSizes: ARRAY[1..numSegs] OF LONGINT;
sample: ARRAY[1..numSamples] OF SampleRecord;
END;

To minimize the disk space required, SampleRecords only keeps track of non-resident segments and won't be written if no non-resident segment had been loaded between the calls to UnloadAllSegments. The new procedure increments DataFile.numSamples and adds an additional SampleRecord to DataFile. The SampleRecord.segmentsLoaded lists all the non-resident segments loaded between calls to UnloadAllSegments.

This data file is very large. The longer a program is tested, the larger the data file becomes. The file can get big enough to make this approach impossible.

Analyzing the data (step 3)

You can use this data to produce a good set of segment mappings. I chose the method outlined here because it is relatively simple and it produces results that are optimal in one category and reasonable in others.

This analysis algorithm gives the absolute minimum necessary code reserve (given that it only creates segment mappings) and reasonable segmentation for minimizing the number of segment loads.

The algorithm works by analyzing samples in order of totalCodeReserveUsage (maximum to minimum). Within each sample segment, combinations are tried (in order of most commonly loaded segments to least commonly loaded segments). If a potential segment mapping does not cause any sample to exceed the peakCodeReserveUsage, it is accepted and the next possible mapping is tried. As a by product, the algorithm can also create the seg! and mem! resources needed to define the temporary memory reserve.

The following pseudo-code shows the algorithm:

FOR sampleNum := 1 TO numSamples DO
BEGIN
    {sort samples from largest code reserve size to smallest}
    SortSamplesByMaxCodeReserveUsageStartingWith(sampleNum);
    sampleToAnalyze := dataFile.sample[sampleNum];
    {Sort segment list in sample by order of }
    { maximum to minimum use}
    SortSegsByMaxUseInSample(sampleToAnalyze);
    FOR mapToSegNum := 1 TO numSegs DO
        BEGIN
        toSegment := sampleToAnalyze.segmentsLoaded[mapToSegNum];
        FOR mapFromSegNum := mapToSegNum + 1 TO numSegs DO
            BEGIN
            fromSegment := 
                sampleToAnalyze.segmentsLoaded[mapFromSegNum];
            {if combining segments doesn't cause any sample to}
            {exceed maxCodeReserve then do it}
            {also could check 32k per segment limit}
            IF CombinedSegmentsWithinMax(toSegment,fromSegment) THEN
                BEGIN
                {create Segment mapping}
                SegmentTogether(toSegment,fromSegment);
                {fix samples as totalCodeReserveUsage etc. may }
                { now be wrong}
                FixDataFileToReflectMapping(toSegment,fromSegment);
                END; {IF}
            END; {FOR mapFromSegNum}
        END; {FOR mapToSegNum}
    END; {FOR sampleNum}

Conclusions

These two schemes are first attempts (by a structural engineer, not a software engineer) to design an automatic segmentation mechanism for MacApp. The statistical analysis approach is limited because it relies on the quality of its empirical rules, but will probably produce reasonable results. The time history approach will produce optimal results (judged by code reserve size) if the history is representative and still small enough that it can be practically stored on disk.

The MacApp team at Apple can surely improve upon these methods, or more likely find a better alternative. MacAppers everywhere hope it's soon.

 

Community Search:
MacTech Search:

Software Updates via MacUpdate

SoundPrism Link Edition (Music)
SoundPrism Link Edition 1.0 Device: iOS Universal Category: Music Price: $4.99, Version: 1.0 (iTunes) Description: ***Introductory price for a the first few days after launch - if you're reading this, get it while it's fresh out of... | Read more »
Pre-register now for hack and slasher An...
Fincon, which won Facebook's Studio to Watch award in 2015, has announced that pre-registration is now open for the massive 3.0 update for its award-winning hack and slasher Angel Stone. Angel Stone is a post-apocalyptic action RPG in which the... | Read more »
Google has named Piano Tiles 2 as its Be...
Google has named Piano Tiles 2, which launched back in August last year, as its Best Game of 2015. If you're yet to play it, now's a good time to do so. It's a sequel to the hugely successful viral hit Piano Tiles (Don't Tap the White Tile) but... | Read more »
Sago Mini Trucks and Diggers (Education...
Sago Mini Trucks and Diggers 1.0 Device: iOS Universal Category: Education Price: $2.99, Version: 1.0 (iTunes) Description: Drive a dump truck with Rosie the hamster! Pile dirt high and dig deep in the ground with diggers, cranes and... | Read more »
Day One 2 Journal + Notes (Lifestyle)
Day One 2 Journal + Notes 2.0 Device: iOS Universal Category: Lifestyle Price: $4.99, Version: 2.0 (iTunes) Description: **New Release Sale—50% off for a Limited Time!** Capture life as you live it. From once-in-a-lifetime events to... | Read more »
Pull My Tongue (Games)
Pull My Tongue 1.0 Device: iOS Universal Category: Games Price: $.99, Version: 1.0 (iTunes) Description: Meet Greg. A hungry little chameleon that loves popcorn! Help feed Greg by pulling his tongue through a series of fun puzzles... | Read more »
Adventures of Mana (Games)
Adventures of Mana 1.0.0 Device: iOS Universal Category: Games Price: $13.99, Version: 1.0.0 (iTunes) Description: Relive the excitement of Final Fantasy Adventure―a timeless classic remastered for a new generation. ■STORY | Read more »
Puzzle Strike (Games)
Puzzle Strike 1.0 Device: iOS Universal Category: Games Price: $3.99, Version: 1.0 (iTunes) Description: Puzzle Strike is a deckbuilding puzzle game played amongst Fantasy Strike characters. It's inspired by puzzle games such as... | Read more »
Tsuro (Entertainment)
Tsuro 1 Device: iOS Universal Category: Entertainment Price: $3.99, Version: 1 (iTunes) Description: The award-wining puzzle strategy game, now for iPhone, iPad and iPod Touch. LAUNCH SALE - 20% OFF!To celebrate our launch, and to... | Read more »
Away ~ Meditation & mindfulness to s...
Away ~ Meditation & mindfulness to sleep, relax, focus, breathe 1.0.2 Device: iOS Universal Category: Healthcare & Fitness Price: $2.99, Version: 1.0.2 (iTunes) Description: Get Away from stress. Enjoy relaxing nature sounds... | Read more »

Price Scanner via MacPrices.net

12-inch 1.2GHz Retina MacBooks on sale for up...
B&H Photo has 12″ 1.2GHz Retina MacBooks on sale for $180 off MSRP. Shipping is free, and B&H charges NY tax only: - 12″ 1.2GHz Gray Retina MacBook: $1499 $100 off MSRP - 12″ 1.2GHz Silver... Read more
12-inch 1.1GHz Gray Retina MacBook on sale fo...
B&H Photo has the 12″ 1.1GHz Gray Retina MacBook on sale for $1199 including free shipping plus NY sales tax only. Their price is $100 off MSRP, and it’s the lowest price available for this model... Read more
Apple now offering full line of Certified Ref...
Apple now has a full line of Certified Refurbished 2015 21″ & 27″ iMacs available for up to $350 off MSRP. Apple’s one-year warranty is standard, and shipping is free. The following models are... Read more
Free GUI Speedometer – The Ultimate Digital D...
Miami, Florida based RMKapps has announced the official release of GUI Speedometer 1.0, their digital dashboard display developed for iOS devices. GUI Speedometer allows users to track their precise... Read more
FutureCalc: Ergonomic iOS Calculator App For...
London, United Kingdom based Independent software developer and entrepreneur, Hovik Melikyan has announced the release and immediate availability of FutureCalc 1.0, his new ergonomic calculator app... Read more
Save up to $600 with Apple refurbished Mac Pr...
Apple has Certified Refurbished Mac Pros available for up to $600 off the cost of new models. An Apple one-year warranty is included with each Mac Pro, and shipping is free. The following... Read more
Apple’s ‘Next Big Thing’ Drought And Mature M...
Apple’s first fiscal quarter 2016 financial results yielded both good news and bad news. In the positive column, Apple made another boatload of money, reporting record quarterly revenue of $75.9... Read more
12-inch 1.1GHz Gray Retina MacBook on sale fo...
Woodwind & Brasswind has the 12″ 1.1GHz Gray Retina MacBook on sale for $1199.99 including free shipping. Their price is $100 off MSRP, and it’s the lowest price available for this model. Read more
Mac minis on sale for $50-$70 off MSRP
B&H Photo has Mac minis on sale for up to $70 off MSRP including free shipping plus NY sales tax only: - 1.4GHz Mac mini: $449.99 $50 off MSRP - 2.6GHz Mac mini: $649.99 $50 off MSRP - 2.8GHz Mac... Read more
Save $140-$180 with Apple refurbished 2015 Ma...
Apple has Certified Refurbished 2015 11″ and 13″ MacBook Airs (the latest models), available for up to $180 off the cost of new models. An Apple one-year warranty is included with each MacBook, and... Read more

Jobs Board

Restaurant Manager - Apple Gilroy Inc./Apple...
…in every aspect of daily operation. WHY YOU'LL LIKE IT: You'll be the Big Apple . You'll solve problems. You'll get to show your ability to handle the stress and Read more
Simply Mac *Apple* Specialist- Service Repa...
Simply Mac is the largest premier retailer of Apple products in the nation. In order to support our growing customer base, we are currently looking for a driven Read more
*Apple* IOS / OS X Level 3 Support Analyst -...
Needed: Apple Support Analyst (OS X / IOS); 50k-60k/yr +…Horizon Technology Partners has an immediate opening for an Apple IOS / OS X LEVEL 3 Support Analyst. Your Read more
*Apple* - Software QA Manager - Apple Inc. (...
Changing the world is all in a day's work at Apple . If you love innovation, here's your chance to make a career of it. You'll work hard. But the job comes with more than Read more
Restaurant Manager - Apple Gilroy Inc./Apple...
…in every aspect of daily operation. WHY YOU'LL LIKE IT: You'll be the Big Apple . You'll solve problems. You'll get to show your ability to handle the stress and Read more
All contents are Copyright 1984-2011 by Xplain Corporation. All rights reserved. Theme designed by Icreon.