Calibration of the Sun Wireless Toolkit

Background
In my last post, I explained how to calibrate the Sun Wireless Toolkit emulator against a given MIDlet market share.

In this post I show provide simple calibration results from the Sun Wireless Tookit 2.2.

Results
The plot shows how the JBenchmark 2 score changes with VM execution speed for an emulated heap of 1MB.


The first thing to note is that, for an emulated speed of 1000 bytecode instructions/second, the JB2 score is only 120. This means that the toolkit can only provide accurate emulation of approximately 45% of the MIDP 2 terminals on the market today.

Next Steps
The graph, above, needs a whole lot more data points to be useful. The results are taken from a single PCs. It is not clear to me whether results differ on different PCs.

Thanks to Doug Wright at uiActive who actually ran the tests - my PC is still rather sick.

Wireless Toolkit Calibration for Real World Market Share

Applied use of Benchmark Results
From my previous posts here and here, it is possible to use a database of MIDP benchmark results to examine how market share varies according to the MIDlet heap memory and performance requirements.

In this post, I explain how to apply this information in order to calibrate the Sun Wireless Tookit against a particular handset market share.

What Problem Are We Solving?
In my experience, developers spend most of their time on a desktop environment, using an IDE like Eclipse with the Eclipse ME plugin to create and debug a MIDlet. Once the basic development and testing is complete, the MIDlet is transferred to a handset for testing.

Issues arise because of the gulf between the emulator performance and real world handset performance. Calibrating the emulator so that it performs in a similar way to the target handset can help to identify and eliminate performance related problems earlier in the development cycle. Simply put, developers become aware of problems during the development stage, rather than at a later test phase.

Sun Wireless Toolkit Support
The Sun Wireless Toolkit provides all of the tools required to adjust the emulator performance. The user can change:

1. VM speed emulation;
2. graphics latency,
   
3. network throughput; and,
4. heap size.
   

Heap is simple enough. However, for performance, there is no simple mapping between the number of bytecodes/millisecond and real world handset performance. Until now, that is.

Calibration How To
Calibration of the emulator is achieved as follows:

1. Download the JBenchmark 2 MIDlet JAD and JAR files from here; or, point your emulator at http://wap.jbenchmark.com/JBenchmark2.jad.
   
2. Decide the MIDlet market share you require.
3. Determine a heap memory and JB2 score (perhaps using the estimation tools I provided here).
4. Set the heap memory using the Wireless toolkit.
5. Check the 'Enable VM Speed Emulation' box.
6. Set the slider to 500.
7. Run the JB2 MIDlet.
8. Watch Mona Lisa, the little jumping guy, the animated 3D knight, and those those MIDP forms that nobody uses.
9. When the benchmark finishes look at the JB2 score.
10. If your KB2 score is greater than your required score, set your emulator slider to 250, if less than your required score, set it to 750.
    
11. Run the MIDlet again and repeat using a binary chop until you have an emulation speed that matches your market share.

The following screenshots are from my N70 showing the start and results page.

Before anyone asks, I used my N70 because my PC CPU fan just parted company with the CPU to which it should be bonded. As a result, I'm using the home MacBook to write this article and I can't actually get the screenshots from the emulator.

What Can Go Wrong?
The Sun Wireless Toolkit only enables VM emulation in the range 100-1000. This range is somewhat arbitrary and seems to have no grounding in real world handset performance. Thus, you may well find that the 1000 mark is too slow for your target handset JB2 score.

Next Steps
Once I get that CPU re-attached, and assuming there is no other long-term damage to my PC, I will update this post with some real mappings between JB2 and emulator bytecode/second scores. I'll then try and get someone else to do the same thing in the hope that there is a simple, globally applicable relationship between JB2 score and emulator settings (wouldn't that be nice!)

I keep saying that I am going to post market share results based on handset brand, and I will. What I want to do is embed an applet in this blog so readers can select any handset brand etc and generate their own chart. I just have to read how to embed an applet in a blog page...

Visualising MIDP Performance Change

Background
Previously, I posted a density plot that demonstrates how MIDlet market share changes as a function of handset heap memory and performance (as measured by the JBenchmark 2 score).

In this post I provide a visualisation showing how market share has changed over time.

Visualisation
The video, below, shows how performance is improving over time.



The video shows how application market share is affected by heap and performance. Each frame is based on one year's worth of data, up to, and including the month shown. For this reason, there is a large amount of noise in the early frames as there were very few MIDP 2 handsets in 2004.

To obtain a very large market share (90+%) of handsets released in 2007, you have to target pretty much the same performance criteria as you did in 2004. In the mid range, handset performance can be seen to be steadily improving.

What Next?
Next, I will run the numbers for Nokia, Motorola, Samsung and Sony Ericsson handsets.

The MIDP Market Share-o-Matic Estimator

Where Were We?
Previously, I looked at MIDP 2 performance and heap memory and provided tools to estimate market share based off either of these measures. I also pointed out problems with any market share estimate based on just one of these measures.

In this post, I provide market share estimates based on both heap and performance criteria.

The Definitive MIDP 2 Market Share Estimator
And here it is...

Using the graph is simple. The X axis gives the heap memory (in this plot, I only show heap to 2.5 MB as it is below this figure where the rate in change in market share is greatest). The Y axis gives the JBenchmark 2 score (a measure of handset performance). The Z axis gives the market share percentile. Percentiles contour lines are plotted for convenience with the contour labels visible to the right of of the Y axis. The small grey dots on the graph show the data points from which the surface was interpolated.

Using the plot you can:

• Determine the addressable market for an existing application;
• Provide specifications for a MIDlet based on a required market share; or,
• Marvel at just how low the specifications are for a MIDlet that is really mass market.

Notable Features
This graph brings into sharp contrast just how important coding for speed and memory efficiency is in the MIDP world if you really want to address a large market share. At the low end, a small increase in MIDlet heap usage causes a significant loss in market share.

The graph is also notable for exactly the opposite reason.

Suppose you agree up front with your developer to live with a target of 768 KB heap and a 100 JB2 score (i.e, you want a market share of 50%). Then, say your developer then tells you that the completed MIDlet actually requires 1MB heap. The graph shows us that this 33% memory overrun will only cost 5% market share. Then, your developer tells you that, ahem, the finished MIDlet only runs on handsets with a JB2 score of 150!

A disaster? Not really. The graph shows us that the combined effect of the 33% memory and 50% performance overrun is a mere 10% reduction in market share.

Another way of saying this is that the rate in change in market share is greatest at low specification handsets. Given what we already know about the bottom 20% of the market (performance and heap), this should not come as a surprise.

Closing Thoughts
This analysis is useful, but needs extending to show similar graphs on a year by year basis. The analysis also needs to be performed for each of the major handset brands. And that is what I hope to provide next time.

A Trip Down MIDP Memory Lane

Changes in MIDP 2 Memory over Time
In this post, I will examine how heap memory has changed in MIDP 2 devices over time. I will also examine the link between heap memory and handset performance as measured by the JBenchmark 2 benchmark. Finally, I will look at why, as developers, we should be concerned with dynamic heap memory in MIDP terminals.

So We Get More Memory in our Handsets, Right?
The following figure shows the maximum heap memory for 609 MIDP 2 terminals with JBenchmark 2 scores over the period January 2004 through July 2007. Note that the memory axis has a log scale in the range 100 KB through 150,000 KB.
As with the performance plots I presented in an earlier post, this graph goes some way to dispelling the myth that all terminals have more memory in 2007 than they did in 2004. In fact, a large number of terminals are still being released into the market with a maximum heap of less than 1 MB. Indeed, it is astonishing that at least two 2007 vintage terminals report as little as 180 KB heap.

MIDP 2 Heap Memory Percentiles
Once again, from a developer perspective, heap memory percentiles are far more valuable than a scatter plot.
Again, our project manager can use this graph to provide his developer team with a handset terminal matching a required market share for the MIDlet under development. For example, there may be a requirement that the MIDlet address 60% of the MIDP 2 handsets released to July 2007, giving the developers 1 MB of heap at a minimum.

Indeed, a manager might extend this to provide a performance specification based on the JB2 percentile plots I presented previously (the JB2 target for a 60% market share is approximately 100).

Unfortunately, life is not that easy...

The Relationship between Performance and Heap Memory in MIDP Terminals
Our project manager, above, is assuming that performance and memory are ranked equally on MIDP 2 terminals.

The plot, below, shows that (unfortunately) there is no relationship between heap memory and performance as measured by the JBenchmark 2 score.
Rather annoyingly, what this says is if a you choose a 60% market share based off heap memory percentiles, your developers can salivate over a nice juicy 1 MB heap, but they then find that they must target a JB2 score of 38!

Unfortunately, things get even worse.

Dynamic Heap in MIDP 2
Many MIDP 2 terminals implement a memory sharing arrangement called dynamic or adaptive heap. This is where the heap is used for other applications as well as the JVM. What it means for a developer is that a clean handset that is used solely for test and development, may have a heap memory that is not representative of the same handset in the field.

The good news is that the folks at JBenchmark record the minimum and maximum heap seen from a handset under test, and these are from a variety of terminals 'in the wild'. Thus, the range in JBenchmark heap reports is likely to be a good indicator of real world terminal conditions. Maximum and minimum heap from 609 JBenchmark 2 results are shown, below.

Of the 609 handsets used in this study, 252 JBenchmark results (41%) show a difference between the high and low heap memory reported by a handset under test. So, as a developer, if you target a handset with 1 MB heap, and your MIDlet is optimised so that it uses every last byte of this space, you can expect no less than 40% of your terminal market to fail in the field as other applications eat into the precious heap via their shared memory architecture.

In some cases, the heap delta is very significant (the plotted point falls well below the 1:1 line). Of course, this will only be an issue when working with lower end terminals; losing 2 MB of a terminal with a maximum of 8 MB is likely to be far less important to the success of a MIDlet than losing 200 KB from a terminal with maximum of 1 MB heap. Dynamic heap is then an issue developers should be concerned about.

Where to From Here?
In my next post I will finally answer some questions concerning market share by presenting a density plot showing the market share than can be expected from a heap / performance pair.

MIDP 2 Performance Frequency Distribution

Background
Previously, I looked at performance percentiles for MIDP 2 handsets over time and said I would provide an analysis of performance by handset vendor. Instead, I will stay at a higher level and look at some more descriptive statistics regarding MIDP 2 performance (JB2 scores) instead over the period January 2004 through July 2007.

The Analysis
The figure below shows the frequency distribution and cumulative frequency distribution of 609 handset JB2 scores.

The cumulative frequency curve approximates the percentile plot I presented in a prior post.

The distribution of JB2 scores has a strong negative skew (γ₁ = 1.07795). That is, more data is concentrated in the left tail than would be expected in a normally distributed data set. The distribution is also 'peaky' with kurtosis (γ₂ = 0.43981). This tells us that, as developers, we should be looking at the median rather than the arithmetic mean when making a decision regarding market share for our MIDlet.

In the JB2 data set, the arithmetic mean of 173.15 JB2 score is significantly greater than the median of 130. If we had naively targeted a score of 173 and expected a 50% market share, we would actually achieve a market share of approximately 35%. (Bear in mind also that the original Motorola RAZR has a JB2 score of only 49.)

Ranking of Release Date and Performance
In order to better visualise just how poor the relationship between release date and MIDP 2 performance is, the rank of the handset release date is plotted against the rank of the corresponding JB2 score. Ranks are normalised in the range 0 through 1.

This plot is another simple way of showing that there is no relationship between handset performance as measured by the JB2 benchmark and release date. The data set has a Kendall tau rank correlation coefficient of τ = 0.196 indicating a poor correlation.

Summary
What this and my previous posts tell us is that as developers we cannot build our mobile application with an expectation that handset performance is doubling every year. If we assume that the 609 handsets with a JBenchmark 2 score that were used in this analysis are representative of the entire population of MIDP 2 terminals, then we must continue to target low performance terminals.

As I said previously, this analysis is simplistic because it does not take into account market share, memory and other considerations. I will publish an analysis of these factors in coming weeks.