Detecting and fixing overflows - math

we have a particle detector hard-wired to use 16-bit and 8-bit buffers. Every now and then, there are certain [predicted] peaks of particle fluxes passing through it; that's okay. What is not okay is that these fluxes usually reach magnitudes above the capacity of the buffers to store them; thus, overflows occur. On a chart, they look like the flux suddenly drops and begins growing again. Can you propose a [mostly] accurate method of detecting points of data suffering from an overflow?
P.S. The detector is physically inaccessible, so fixing it the 'right way' by replacing the buffers doesn't seem to be an option.
Update: Some clarifications as requested. We use python at the data processing facility; the technology used in the detector itself is pretty obscure (treat it as if it was developed by a completely unrelated third party), but it is definitely unsophisticated, i.e. not running a 'real' OS, just some low-level stuff to record the detector readings and to respond to remote commands like power cycle. Memory corruption and other problems are not an issue right now. The overflows occur simply because the designer of the detector used 16-bit buffers for counting the particle flux, and sometimes the flux exceeds 65535 particles per second.
Update 2: As several readers have pointed out, the intended solution would have something to do with analyzing the flux profile to detect sharp declines (e.g. by an order of magnitude) in an attempt to separate them from normal fluctuations. Another problem arises: can restorations (points where the original flux drops below the overflowing level) be detected by simply running the correction program against the reverted (by the x axis) flux profile?

int32[] unwrap(int16[] x)
{
// this is pseudocode
int32[] y = new int32[x.length];
y[0] = x[0];
for (i = 1:x.length-1)
{
y[i] = y[i-1] + sign_extend(x[i]-x[i-1]);
// works fine as long as the "real" value of x[i] and x[i-1]
// differ by less than 1/2 of the span of allowable values
// of x's storage type (=32768 in the case of int16)
// Otherwise there is ambiguity.
}
return y;
}
int32 sign_extend(int16 x)
{
return (int32)x; // works properly in Java and in most C compilers
}
// exercise for the reader to write similar code to unwrap 8-bit arrays
// to a 16-bit or 32-bit array

Of course, ideally you'd fix the detector software to max out at 65535 to prevent wraparound of the sort that is causing your grief. I understand that this isn't always possible, or at least isn't always possible to do quickly.
When the particle flux exceeds 65535, does it do so quickly, or does the flux gradually increase and then gradually decrease? This makes a difference in what algorithm you might use to detect this. For example, if the flux goes up slowly enough:
true flux measurement
5000 5000
10000 10000
30000 30000
50000 50000
70000 4465
90000 24465
60000 60000
30000 30000
10000 10000
then you'll tend to have a large negative drop at times when you have overflowed. A much larger negative drop than you'll have at any other time. This can serve as a signal that you've overflowed. To find the end of the overflow time period, you could look for a large jump to a value not too far from 65535.
All of this depends on the maximum true flux that is possible and on how rapidly the flux rises and falls. For example, is it possible to get more than 128k counts in one measurement period? Is it possible for one measurement to be 5000 and the next measurement to be 50000? If the data is not well-behaved enough, you may be able to make only statistical judgment about when you have overflowed.

Your question needs to provide more information about your implementation - what language/framework are you using?
Data overflows in software (which is what I think you're talking about) are bad practice and should be avoided. While you are seeing (strange data output) is only one side effect that is possible when experiencing data overflows, but it is merely the tip of the iceberg of the sorts of issues you can see.
You could quite easily experience more serious issues like memory corruption, which can cause programs to crash loudly, or worse, obscurely.
Is there any validation you can do to prevent the overflows from occurring in the first place?

I really don't think you can fix it without fixing the underlying buffers. How are you supposed to tell the difference between the sequences of values (0, 1, 2, 1, 0) and (0, 1, 65538, 1, 0)? You can't.

How about using an HMM where the hidden state is whether you are in an overflow and the emissions are observed particle flux?
The tricky part would be coming up with the probability models for the transitions (which will basically encode the time-scale of peaks) and for the emissions (which you can build if you know how the flux behaves and how overflow affects measurement). These are domain-specific questions, so there probably aren't ready-made solutions out there.
But one you have the model, everything else---fitting your data, quantifying uncertainty, simulation, etc.---is routine.

You can only do this if the actual jumps between successive values are much smaller than 65536. Otherwise, an overflow-induced valley artifact is indistinguishable from a real valley, you can only guess. You can try to match overflows to corresponding restorations, by simultaneously analysing a signal from the right and the left (assuming that there is a recognizable base line).
Other than that, all you can do is to adjust your experiment by repeating it with different original particle flows, so that real valleys will not move, but artifact ones move to the point of overflow.

Related

Statistical best fit for gesture detection

I have a linear regression equation from school , which gives a value between 1 and -1 indicative of whether or not a set of data points are close enough to a linear function
and the equation given here
http://people.hofstra.edu/stefan_waner/realworld/calctopic1/regression.html
under best fit of a line. I would like to use these to do simple gesture detection based on a point in 3-space (x,y,z) - forward, back, left, right, up, down. First I would see if they fall on a line in 2 of the 3 dimensions, then I would see if that line's slope approached zero or infinity.
Is this fast enough for functional gesture recognition? If not, could someone propose an alternative algorithm?
If I've understood your question correctly then (1) the calculation you describe here would probably be plenty fast enough, (2) it may not actually do what you want, and (3) the stuff that'll be slow in an actual implementation would lie elsewhere.
So, I think you're proposing to do this. (1) Identify the positions of ... something ... (the user's hand, perhaps) in three-dimensional space, at several successive times. (2) For (say) each of {x,y} and {x,z}, look at those two coordinates of each point, compute the correlation coefficient (which is what your formula describes) and see whether it's close to +-1. (3) If both correlation coefficients are close to +-1 then the points lie approximately on a straight line; calculate the gradient of that line (using a formula similar to that of the correlation coefficient). (4) If the gradients are both very close to 0 or +- infinity, then your line is approximately parallel to one axis, which is the case you're trying to recognize.
1: Is it fast enough? You might perhaps be sampling at 50 frames per second or thereabouts, and your gestures might take a second to execute. So you'll have somewhere on the order of 50 positions. So, the total number of arithmetic operations you'll need is maybe a few hundred (including a modest number of square roots). In the worst case, you might be doing this in emulated floating-point on a slow ARM processor or something; in that case, each arithmetic operation might take a couple of hundred cycles, so the whole thing might be 100k cycles, which for a really slow processor running at 100MHz would be about a millisecond. You're not going to have any problem with the time taken to do this calculation.
2: Is it the right thing? It's not clear that it's the right calculation. For instance, suppose your user's hand moves back and forth rapidly several times along the x-axis; that will give you a positive result; is that what you want? Suppose the user attempts the gesture you want but moves at slightly the wrong angle; you may get a negative result. Suppose they move exactly along the x-axis for a bit and then along the y-axis for a bit; then the projections onto the {x,y}, {x,z} and {y,z} planes will all pass your test. These all seem like results you might not want.
3: Is it where the real cost will lie? This all assumes you've already got (x,y,z) coordinates. Getting those is probably going to be more expensive than processing them. For instance, if you have a camera-based system of some kind then there'll be some nontrivial image processing for every frame. Or perhaps you're integrating up data from accelerometers (which, by the way, is likely to give nasty inaccurate position results); the chances are that you're doing some filtering and other calculations to get position data. I bet that the cost of performing a calculation like this one will be substantially less than the cost of getting the coordinates in the first place.

OpenCL computation times much longer than CPU alternative

I'm taking my first steps in OpenCL (and CUDA) for my internship. All nice and well, I now have working OpenCL code, but the computation times are way too high, I think. My guess is that I'm doing too much I/O, but I don't know where that could be.
The code is for the main: http://pastebin.com/i4A6kPfn, and for the kernel: http://pastebin.com/Wefrqifh I'm starting to measure time after segmentPunten(segmentArray, begin, eind); has returned, and I end measuring time after the last clEnqueueReadBuffer.
Computation time on a Nvidia GT440 is 38.6 seconds, on a GT555M 35.5, on a Athlon II X4 5.6 seconds, and on a Intel P8600 6 seconds.
Can someone explain this to me? Why are the computation times are so high, and what solutions are there for this?
What is it supposed to do: (short version) to calculate how much noiseload there is made by an airplane that is passing by.
long version: there are several Observer Points (OP) wich are the points in wich sound is measured from an airplane thas is passing by. The flightpath is being segmented in 10.000 segments, this is done at the function segmentPunten. The double for loop in the main gives OPs a coordinate. There are two kernels. The first one calculates the distance from a single OP to a single segment. This is then saved in the array "afstanden". The second kernel calculates the sound load in an OP, from all the segments.
Just eyeballing your kernel, I see this:
kernel void SEL(global const float *afstanden, global double *totaalSEL,
const int aantalSegmenten)
{
// ...
for(i = 0; i < aantalSegmenten; i++) {
double distance = afstanden[threadID * aantalSegmenten + i];
// ...
}
// ...
}
It looks like aantalSegmenten is being set to 1000. You have a loop in each
kernel that accesses global memory 1000 times. Without crawling though the code,
I'm guessing that many of these accesses overlap when considering your
computation as a whole. It this the case? Will two work items access the same
global memory? If this is the case, you will see a potentially huge win on the
GPU from rewriting your algorithm to partition the work such that you can read
from a specific global memory only once, saving it in local memory. After that,
each work item in the work group that needs that location can read it quickly.
As an aside, the CL specification allows you to omit the leading __ from CL
keywords like global and kernel. I don't think many newcomers to CL realize
that.
Before optimizing further, you should first get an understanding of what is taking all that time. Is it the kernel compiles, data transfer, or actual kernel execution?
As mentioned above, you can get rid of the kernel compiles by caching the results. I believe some OpenCL implementations (the Apple one at least) already do this automatically. With other, you may need to do the caching manually. Here's instructions for the caching.
If the performance bottle neck is the kernel itself, you can probably get a major speed-up by organizing the 'afstanden' array lookups differently. Currently when a block of threads performs a read from the memory, the addresses are spread out through the memory, which is a real killer for GPU performance. You'd ideally want to index array with something like afstanden[ndx*NUM_THREADS + threadID], which would make accesses from a work group to load a contiguous block of memory. This is much faster than the current, essentially random, memory lookup.
First of all you are measuring not the computation time but the whole kernel read-in/compile/execute mumbo-jumbo. To do a fair comparison measure the computation time from the first "non-static" part of your program. (For example from between the first clSetKernelArgs to the last clEnqueueReadBuffer.)
If the execution time is still too high, then you can use some kind of profiler (such as VisualProfiler from NVidia), and read the OpenCL Best Practices guid which is included in the CUDA Toolkit documentation.
To the raw kernel execution time: Consider (and measure) that do you really need the double precision for your calculation, because the double precision calculations are artificially slowed down on the consumer grade NVidia cards.

OpenCL image histogram

I'm trying to write a histogram kernel in OpenCL to compute 256 bin R, G, and B histograms of an RGBA32F input image. My kernel looks like this:
const sampler_t mSampler = CLK_NORMALIZED_COORDS_FALSE |
CLK_ADDRESS_CLAMP|
CLK_FILTER_NEAREST;
__kernel void computeHistogram(read_only image2d_t input, __global int* rOutput,
__global int* gOutput, __global int* bOutput)
{
int2 coords = {get_global_id(0), get_global_id(1)};
float4 sample = read_imagef(input, mSampler, coords);
uchar rbin = floor(sample.x * 255.0f);
uchar gbin = floor(sample.y * 255.0f);
uchar bbin = floor(sample.z * 255.0f);
rOutput[rbin]++;
gOutput[gbin]++;
bOutput[bbin]++;
}
When I run it on an 2100 x 894 image (1,877,400 pixels) i tend to only see in or around 1,870,000 total values being recorded when I sum up the histogram values for each channel. It's also a different number each time. I did expect this since once in a while two kernels probably grab the same value from the output array and increment it, effectively cancelling out one increment operation (I'm assuming?).
The 1,870,000 output is for a {1,1} workgroup size (which is what seems to get set by default if I don't specify otherwise). If I force a larger workgroup size like {10,6}, I get a drastically smaller sum in my histogram (proportional to the change in workgroup size). This seemed strange to me, but I'm guessing what happens is that all of the work items in the group increment the output array value at the same time, and so it just counts as a single increment?
Anyways, I've read in the spec that OpenCL has no global memory syncronization, only syncronization within local workgroups using their __local memory. The histogram example by nVidia breaks up the histogram workload into a bunch of subproblems of a specific size, computes their partial histograms, then merges the results into a single histogram after. This doesn't seem like it'll work all that well for images of arbitrary size. I suppose I could pad the image data out with dummy values...
Being new to OpenCL, I guess I'm wondering if there's a more straightforward way to do this (since it seems like it should be a relatively straightforward GPGPU problem).
Thanks!
As stated before, you write into a shared memory unsynchronized and non atomic. This leads to errors. If the picture is big enough, I have a suggestion:
Split your work group into a one dimensional one for cols or rows. Use each kernel to sum up the histogram for the col or row and afterwards sum it globally with atomic atom_inc. This brings the most sum ups in private memory which is much faster and reduces atomic ops.
If you work in two dimensions you can do it on parts of the picture.
[EDIT:]
I think, I have a better answer: ;-)
Have a look to: http://developer.download.nvidia.com/compute/opencl/sdk/website/samples.html#oclHistogram
They have an interesting implementation there...
Yes, you're writing to a shared memory from many work-items at the same time, so you will lose elements if you don't do the updates in a safe way (or worse ? Just don't do it). The increase in group size actually increases the utilization of your compute device, which in turn increases the likelihood of conflicts. So you end up losing more updates.
However, you seem to be confusing synchronization (ordering thread execution order) and shared memory updates (which typically require either atomic operations, or code synchronization and memory barriers, to make sure the memory updates are visible to other threads that are synchronized).
the synchronization+barrier is not particularly useful for your case (and as you noted is not available for global synchronization anyways. Reason is, 2 thread-groups may never run concurrently so trying to synchronize them is nonsensical). It's typically used when all threads start working on generating a common data-set, and then all start to consume that data-set with a different access pattern.
In your case, you can use atomic operations (e.g. atom_inc, see http://www.cmsoft.com.br/index.php?option=com_content&view=category&layout=blog&id=113&Itemid=168). However, note that updating a highly contended memory address (say, because you have thousands of threads trying all to write to only 256 ints) is likely to yield poor performance. All the hoops typical histogram code goes through are there to reduce the contention on the histogram data.
You can check
The histogram example from AMD Accelerated Parallel Processing (APP) SDK.
Chapter 14 - Image Histogram of OpenCL Programming Guide book (ISBN-13: 978-0-321-74964-2).
GPU Histogram - Sample code from Apple

Finding area of straight line with graph (Math question but needed for flot)

Okay, so this is a straight math question and I read up on meta that those need to be written to sound like programming questions. I'll do my best...
So I have graph made in flot that shows the network usage (in bytes/sec) for the user. The data is 4 minutes apart when there is activity, and otherwise set at the start of the usage range (let's say day 1) and the end of the range (day 7). The data is coming from a CGI script I have no control over, so I'm fairly limited in what I can provide the user.
I never took trig or calculus, so I'm pretty much in over my head. What I want is for the user to have the option to click any point on the graph and see their bandwidth usage for that moment. Since the lines between real data points are drawn straight, this can be done by getting the points before and after where the user has clicked and finding the y-interval.
It took me weeks to finally get a helpful math person to explain this to me. Everyone else has insisted on trying to teach me Riemann sum techniques and all sorts of other heavy stuff that not only is confusing to me, doesn't seem necessary for the problem.
But I also want the user to be able to highlight the graph from two arbitrary points on the y-axis (time) to get the amount of network usage total during that range. I know this would be inaccurate, but I need it to be the right inaccurate using a solid equation.
I thought this was the area under the line, but experiments with much simpler graphs makes this seem just far too high. I figured out I could take the distance from y2 - y1 and multiply it by x2 - x1 and then divide by two to get the area of the graph below the line like a triangle, but again, the numbers seemed to high. (maybe they are just big numbers and I don't get this math stuff at all).
So what I need, if anyone would be really awesome enough to provide it before this question is closed down for being too pure-math, is either the name of the concept I should be researching or the equation itself. Or the bad news that I do need advanced math to get an accurate result.
I am not bad at math, just as a last note, I just am not familiar with math beyond 10th grade and so I need some place to start. All the math sites seem to keep it too simple or way over my paygrade.
If I understood correctly what you're asking (and that is somewhat doubtful), you should find what you seek in these links:
Linear interpolation
(calculating the value of the point in between)
Trapezoidal rule
(calculating the area below the "curve")
*****Edit, so we can get this over :) without much ado:*****
So I have graph made in flot that shows the network usage (in bytes/sec) for the user. The data is 4 minutes apart when there is activity, and otherwise set at the start of the usage range (let's say day 1) and the end of the range (day 7). The data is coming from a CGI script I have no control over, so I'm fairly limited in what I can provide the user.
What is a "flot" ?
Okey, so you have speed on y axis [in bytes/sec]; and time on x axis in [sec], right?
That means, that if you're flotting (I'm bored, yes :) speed over time, in linear segments, interpolating at some particular point in time you'll get speed at that particular point in time.
If you wish to calculate how much bandwidth you've spend, you need to determine the area beneath that curve. The area from point "a" to point "b" will determine the spended bandwidth in [bytes] in that time period.
It took me weeks to finally get a helpful math person to explain this to me. Everyone else has insisted on trying to teach me Riemann sum techniques and all sorts of other heavy stuff that not only is confusing to me, doesn't seem necessary for the problem.
In the immortal words of Snoopy: "Good grief !"
But I also want the user to be able to highlight the graph from two arbitrary points on the y-axis (time) to get the amount of network usage total during that range. I know this would be inaccurate, but I need it to be the right inaccurate using a solid equation.
It would not be inaccurate.
It would be actually perfectly accurate (well, apart from roundoff error in bytes :), since you're using linear interpolation on linear segments.
I thought this was the area under the line, but experiments with much simpler graphs makes this seem just far too high. I figured out I could take the distance from y2 - y1 and multiply it by x2 - x1 and then divide by two to get the area of the graph below the line like a triangle, but again, the numbers seemed to high. (maybe they are just big numbers and I don't get this math stuff at all).
"like a triangle" --> should be "like a trapezoid"
If you do deltax*(y2-y1)/2 you will get the area, yes (this works only for linear segments). This is the basis principle of trapezoidal rule.
If you're uncertain about what you're calculating use dimensional analysis: speed is in bytes/sec, time is in sec, bandwidth is in bytes. Multiplying speed*time=bandwidth, and so on.
What I want is for the user to have
the option to click any point on the
graph and see their bandwidth usage
for that moment. Since the lines
between real data points are drawn
straight, this can be done by getting
the points before and after where the
user has clicked and finding the
y-interval.
Yes, that's a good way to find that instantaneous value. When you report that value back, it's in the same units as the y-axis, so that means bytes/sec, right?
I don't know how rapidly the rate changes between points, but it's even simpler if you simply pick the closest point and report its value. You simplify your problem without sacrificing too much accuracy.
I thought this was the area under the
line, but experiments with much
simpler graphs makes this seem just
far too high. I figured out I could
take the distance from y2 - y1 and
multiply it by x2 - x1 and then divide
by two to get the area of the graph
below the line like a triangle, but
again, the numbers seemed to high.
(maybe they are just big numbers and I
don't get this math stuff at all).
To calculate the total bytes over a given time interval, you should find the index closest to the starting and ending point and multiply the value of y by the spacing of your x-points and add them all together. That will give you the total # of bytes consumed during that time interval, but there's one more wrinkle you might have forgotten.
You said that the points come in "4 minutes apart", and your y-axis is in bytes/second. Remember that units matter. Your area is the sum of bytes/second times a spacing in minutes. To make the units come out right you have to multiply by 60 seconds/minute to get the final value of bytes that you want.
If that "too high" value is still off, consider units again. It's 1024 bytes per kbyte, and 1024*1024 bytes per MB. Check the units of the values you're checking the calculation against.
UPDATE:
No wonder you're having problems. Your original question CLEARLY stated bytes/sec. Even this question is imprecise and confusing. How did you arrive at "amount of data" at a given time stamp? Are those the total bits transferred since the last time stamp? If yes, simply add the values between the start and end of the interval you want and convert to the units convenient for you.
The network usage total is not in bytes (kilo-, mega-, whatever) per second. It would be in just straight bytes (or kilo-, or whatever).
For example, 2 megabytes per second over an interval of 10 seconds would be 20 megabytes total. It would not be 20 megabytes per second.
Or do you perhaps want average bytes per second over an interval?
This would be a lot easier for you if you would accept that there is well-established terminology for the concepts that you are having trouble expressing concisely or accurately, and that these mathematical terms have been around far longer than you. Since you've clearly gone through most of the trouble of understanding the concepts, you might as well break down and start calling them by their proper names.
That said:
There are 2 obvious ways to graph bandwidth, and two ways you might be getting the bandwidth data from the server. First, there's the cumulative usage function, which for any time is simply the total amount of data transferred since the start of the measurement. If you plot this function, you get a graph that never decreases (since you can't un-download something). The units of the values of this function will be bytes or kB or something like that.
What users are typically interested is in the instantaneous usage function, which is an indicator of how much bandwidth you are using right now. This is what users typically want to see. In mathematical terms, this is the derivative of the cumulative function. This derivative can take on any value from 0 (you aren't downloading) to the rated speed of your network link (indicating that you're pushing as much data as possible through your connection). The units of this function are bytes per second, or something related like Mbps (megabits per second).
You can approximate the instantaneous bandwidth with the average data usage over the past few seconds. This is computed as
(number of bytes transferred)
-----------------------------------------------------------------
(number of seconds that elapsed while transferring those bytes)
Generally speaking, the smaller the time interval, the more accurate the approximation. For simplicity's sake, you usually want to compute this as "number of bytes transferred since last report" divided by "number of seconds since last report".
As an example, if the server is giving you a report every 4 minutes of "total number of bytes transferred today", then it is giving you the cumulative function and you need to approximate the derivative. The instantaneous bandwidth usage rate you can report to users is:
(total transferred as of now) - (total as of 4 minutes ago) bytes
-----------------------------------------------------------
4*60 seconds
If the server is giving you reports of the form "number of bytes transferred since last report", then you can directly report this to users and plot that data relative to time. On the other hand, if the user (or you) is concerned about a quota on total bytes transferred per day, then you will need to transform the (approximately) instantaneous data you have into the cumulative data. This process, known as computing the integral, is the opposite of computing the derivative, and is in some ways conceptually simpler. If you've kept track of each of the reports from the server and the timestamp, then for each time, the value you plot is the total of all the reports that came in before that time. If you're doing this in realtime, then every time you get a new report, the graph jumps up by the amount in that report.
I am not bad at math, ... I just am not familiar with math beyond 10th grade
This is like saying "I'm not bad at programming, I have no trouble with ifs and loops but I never got around to writing more than one function."
I would suggest you enrol in a maths class of some kind. An understanding of matrices and the basics of calculus gives you an appreciation of many things, and can be useful in all sorts of areas. You'll be able to understand more of Wikipedia articles and SO answers - and questions!
If you can't afford that, try to find some lecture videos or something.
Everyone else has insisted on trying to teach me Riemann sum techniques
I can't see why. You don't need them for this - though if you had learned them, I expect you would find it easier to come up with a solution. You see, Riemann sums attempt to give you a "familiar" notion of area. The sort of area you (hopefully) learned years ago.
Getting the area below your usage graph between two points will tell you (approximately) how much was used over that period.
How do you find the area of a floor plan? You break it up into rectangles and triangles, find the area of each, and add them together. You can do the same thing with your graph, basically. Someone has worked out a simple way of doing this called the trapezoidal rule. It's just a matter of choosing how to divide your graph into strips, and in your case this is easy: just use the data points themselves as dividers. (You'll also need to work out the value of the graph at the left and right ends of the region selected by the user, using linear interpolation.)
If there's anything I've said that isn't clear to you (as there may well be), please leave a comment.

Using an epsilon value to determine if a ball in a game is not moving?

I have balls bouncing around and each time they collide their speed vector is reduced by the Coefficient of Restitution.
Right now my balls CoR for my balls is .80 . So after many bounces my balls have "stopped" rolling because their speed has becoming some ridiculously small number.
In what stage is it appropriate to check if a speed value is small enough to simply call it zero (so I don't have the crazy jittering of the balls reacting to their micro-velocities). I've read on some forums before that people will sometimes use an epsilon constant, some small number and check against that.
Should I define an epsilon constant and do something like:
if Math.abs(velocity.x) < epsilon then velocity.x = 0
Each time I update the balls velocity and position? Is this what is generally done? Would it be reasonable to place that in my Vector classes setters for x and y? Or should I do it outside of my vector class when I'm calculating the velocities.
Also, what would be a reasonable epsilon value if I was using floats for my speed vector?
A reasonable value for epsilon is going to depend on the constraints of your system. If you are representing the ball graphically, then your epsilon might correspond to, say, a velocity of .1 pixels a second (ensuring that your notion of stopping matches the user's experience of the screen objects stopping). If you're doing a physics simulation, you'll want to tune it to the accuracy to which you're trying to measure your system.
As for how often you check - that depends as well. If you're simulating something in real time, the extra check might be costly, and you'll want to check every 10 updates or once per second or something. Or performance might not be an issue, and you can check with every update.
Instead of an epsilon for an IsStillMoving function, maybe you could use an UpdatePosition function, scheduled on an object-by-object basis based on its velocity.
I'd do something like this (in my own make-it-up-as-you-go pseudocode):
void UpdatePosition(Ball b) {
TimeStamp now = Clock.GetTime();
float secondsSinceLastUpdate = now.TimeSince(b.LastUpdate).InSeconds;
Point3D oldPosition = b.Position;
Point3D newPosition = CalculatePosition(b.Position, b.Velocity, interval);
b.MoveTo(newPosition);
float epsilonOfAccuracy = 0.5; // Accurate to one half-pixel
float pixelDistance = Camera.PixelDistance(oldPosition, newPosition);
float fps = System.CurrentFramesPerSecond;
float secondsToMoveOnePixel = (pixelDistance * secondsSinceLastUpdate) / fps;
float nextUpdateInterval = secondsToMoveOnePixel / epsilonOfAccuracy;
b.SetNextUpdateAt(now + nextUpdateInterval);
}
Balls moving very quickly would get updated on every frame. Balls moving more slowly might update every five or ten frames. And balls that have stopped (or nearly stopped) would update only very very rarely.
IMO your epsilon approach is fine. I would just experiment to see what looks or feels natural to the animation in the game.
Epsilon by nature is the smallest possible increment. Unfortunately, computers have different "minimal" increments of their own depending on the floating point representation. I would be very careful (and might even go higher than what I would calculate just for safety) playing around with that, especially if I want a code to be portable.
You may want to write a function that figures out the minimal increment on your floats rather than use a magic value.

Resources