Marat Fairuzov, B626 project 4

Marat Fairuzov
B629, Project 4
A distributed filesystem synchronizer using Java RMI

Part 2: File synchronizer implementation improvements

1. General ideas

The original implementation was (in my personal opinion) a good illustration of the RMI usage, but generally unsuited for a practical application, not only in terms of its performance, but also in terms of the underlying philosophy. The usage of numerous classes and methods in many cases was not justified by the needs of the problem. During the execution of the synchronizer the file systems remain static, they are not being modified. So there is no need for the dynamic communication between the two parts of the synchronizer.
Maybe another language (Perl for example) would be a better choice for this kind of application.
My vision of a good implementation is the following:
1. One of the machines is selected as client. It will do most of the work, so it must be the fastest of the two. The other machine is server. Note, that the client might actually be the remote machine.
2. There is only one remote object on the server; this object is responsible for creating and sending to the client an image of the remote file system, and for receiving and performing a task list from the client. An image of the file system includes the list of files along with their attributes (file or directory, signatures for the files, dirty bits, modification times, etc).
3. The client creates an image of its own file system and gets the image of the remote file system. This might be parallelized, but only 1 additional thread is needed: to send the request to the server and receive the reply, while the main thread is busy computing the image.
4. All the work is done on the images of the file systems. As the client scans through them, it decides what should be done on its side and on the server side and records that information in the task lists.
5. Ideally, the recursion should also be eliminated, i.e. the scan through the file system images should be sequential, because the recursion is an expensive technique, especially in an interpreted program.
6. In the end the client sends server's task list to the server and starts to execute its own tasks. This can also be parallelized, but again, only 1 additional thread is needed, so that the client doesn't have to wait for the server to finish the tasks before it starts on its own.
7. To execute the task list, the server might need to transfer some files from the the client, so the client code should include some methods for that. Ideally, the server should issue just one request for all the files that it needs (again, it can be done in a separate thread).
This design minimizes the number of messages, not necessarily the total amount of the data transfered. It is a well-known fact, that sending one large message is cheaper than sending a number of short messages, even if the total size of the short messages is less that the size of the large message (within certain limits, of course). In this particular problem, with a careful choice of the information sent, this design gives a significant overall improvement. This design also ensures that the communications and I/O operations are clustered, so that the system layers have a chance to optimize the data transmittion.

2. Implementation improvements
Unfortunately, the time constraints and the deficiencies of the original design didn't allow to fully implement the design described above. Although I tried to get as close as possible...

The following bug was fixed: when called with one root path pointing to a non-existent file (and the other root path pointing to a real file) the original version would halt with a NullPointerException.
The NetFile interface was eliminated along with the class that implements it. Some of the methods were re-implemented in the FileSystemImpl class with the necessary modifications. Now all files are referred to by the pair (FileSystem, filename).
The main() method of the Snc class uses the (new) FileSystemImpl.getDirtyTable() method to get images of the local and remote file systems. An image is just a hash table, where each file is inserted with its path as the key, and the following information:
```
    boolean dirty;    boolean hasDirtyDescendant;
    String Path;      boolean Exists;
    boolean IsFile;   boolean IsDirectory;
    String[] List;    // directories only
    Signature Sign;   // files only
```
The Snc.snc() method uses these images to compare the file systems. Instead of calling the remote method for each file (either to do something with that file, or to mark it synchronized), it inserts the precessed files (sometimes with their infos) into the task lists: the list of files to mark synchronized, and the list of files to copy from the other file system, in the form of the array of file infos for these files, and the list of the files.
After Snc.snc finishes, the main() method calls FileSystemImpl.sncFinalize(), passing it the task lists and the source file system reference.
The FileSystemImpl.sncFinalize(), method performs the tasks determined by its arguments. Note, that now all the markSynchronized(), calls are local.
The Snc.snc method still uses the recursion to descend down the directory structure. There are no extra threads in the version that I'm turning in.
There still is a number of rather obvious optimizations that can be performed. But most of them would take a significant amount of time and require debugging and testing.

3. Timing results
The following programs have been compared:

the original implementation (table: Original);
prof.Pierce's improved version (table: Pierce);
the version for which I am turning in the source code (table: Mod13);
an extra version for which I am not turning the source code in (table: Extra);

The first presented set of the measurements was taken at school.cs.indiana.edu at the busiest time (4 - 6 pm). As far as I could tell, there were no CPU-intensive processes running there during my experiments, but the machine was a bit slow. This is the graph, showing each program's time as the percentage of the original implementation:
[school_trans.gif]

The programs ran in the batch mode, but later I had to repeat some of the runs for the following reasons: during one of the runs the original program somehow decided that there was an inconsistency in the file systems and asked for the user decision (I noticed that only a half an hour later); prof.Pierce's improved version once "hung", and I had to kill it; each program once or twice had a time that was obviously a "glitch" (I re-run only the cases that had time around twice the average). The programs didn't run in the order presented in the tables. I was "mixing" them in the following way: after 4 runs of one program another one was run for 4 times and so on. Time is given in seconds, the leftmost column is the run number, two last rows give the average and the standard deviation.

Original
      everyting out of date   everyting up to date
       total    user    sys    total   user    sys
--------------------------------------------------
   1  141.04   22.55   6.93    55.96   4.65   2.42
   2  134.38   23.35   7.53    68.77   5.05   2.64
   3  143.74   24.06   7.34    67.61   4.53   2.69
   4  124.95   22.14   7.42    67.49   4.72   2.57
   5  120.29   21.99   6.95    61.69   4.70   2.59
   6  122.65   23.15   6.86    55.33   4.36   2.43
   7  123.19   22.37   7.19    50.88   4.53   2.44
   8  137.88   21.79   7.18    51.07   4.22   2.36
   9  140.73   22.18   7.48    55.94   4.73   2.86
  10  138.46   22.63   7.60    58.41   4.48   2.31
  11  132.68   22.86   7.34    67.37   4.85   2.23
  12  141.30   23.35   7.36    69.29   4.69   2.27
--------------------------------------------------
avrg  133.44   22.70   7.26    60.82   4.63   2.48
sdev    8.12    0.64   0.23     6.75   0.21   0.18


Pierce
      everyting out of date   everyting up to date
       total    user    sys    total   user    sys
--------------------------------------------------
   1   83.12   14.38   6.28    40.36   7.70   5.08
   2   62.52   15.49   6.60    37.77   6.91   4.41
   3   68.27   15.64   6.58    42.35   7.13   4.93
   4   70.81   14.65   6.49    36.28   6.99   4.83
   5   85.02   15.05   6.35    37.93   6.83   4.56
   6   62.26   14.70   6.48    33.69   6.40   4.05
   7   54.01   13.49   5.93    31.96   6.62   4.48
   8   63.99   15.01   6.24    35.43   6.60   4.45
   9   76.37   14.79   6.80    68.33   8.34   4.91
  10   78.15   15.28   6.86    43.39   7.51   5.47
  11   66.83   13.98   6.63    56.14   8.21   5.48
  12   98.89   16.05   7.62    52.62   7.69   5.56
--------------------------------------------------
avrg   72.52   14.88   6.57    43.02   7.24   4.85
sdev   11.85    0.68   0.40    10.32   0.61   0.46


Mod13
      everyting out of date   everyting up to date
       total    user    sys    total   user    sys
--------------------------------------------------
   1   50.80    6.06   2.35    48.07   6.56   2.49
   2   45.60    5.64   1.97    49.85   6.40   2.66
   3   42.24    5.52   2.02    48.38   6.24   2.37
   4   45.08    5.77   1.81    45.88   6.58   2.21
   5   67.58    5.68   2.35    84.00   7.01   2.64
   6   62.59    6.15   2.07    83.82   6.85   2.54
   7   52.26    6.01   2.11    59.83   6.33   2.70
   8   55.94    5.91   2.25    56.06   6.68   2.70
   9   65.93    6.08   2.25    83.40   6.60   2.70
  10   63.77    5.68   2.21    83.12   6.79   2.93
  11   62.11    5.96   2.50    74.89   7.00   2.73
  12   56.71    5.76   2.29    85.32   6.90   2.65
--------------------------------------------------
avrg   55.88    5.85   2.18    66.89   6.66   2.61
sdev    8.33    0.19   0.18    16.11   0.25   0.18


Extra
      everyting out of date   everyting up to date
       total    user    sys    total   user    sys
--------------------------------------------------
   1   46.69    3.85   1.62    49.87   4.50   1.58
   2   55.85    3.87   1.48    44.30   4.38   1.54
   3   44.07    3.80   1.43    46.88   4.40   1.56
   4   50.90    3.65   1.74    50.58   4.45   1.73
   5   56.94    3.86   1.85    59.15   4.33   2.13
   6   59.64    3.70   1.76    72.63   4.46   2.27
   7   56.78    3.93   1.78    54.45   4.25   1.99
   8   61.35    4.07   1.74    71.50   4.83   2.02
   9   56.80    3.71   1.65    56.55   4.45   2.01
  10   52.89    4.31   1.68    60.08   4.57   1.86
  11   56.06    3.81   1.51    81.62   4.79   1.84
  12   61.25    4.28   2.08    57.28   4.35   1.87
--------------------------------------------------
avrg   54.93    3.90   1.69    58.74   4.48   1.87
sdev    5.19    0.20   0.17    10.80   0.17   0.22

The second presented set of the measurements was taken at guitar.cs.indiana.edu when the machine was idle (4 - 6 am). Nobody was logged in, and except for the processes owned by root and me, there was only a couple of "abandoned" java's in the sleep state. Therefore I tend to think that this set reflects the relative times better than the previous. This is the percentage graph:
[guitar_trans.gif]

These runs went without large time variations, except for #1 in Pierce and #4 in Mod13, which apparently were due to some system daemons. None of the experiments was re-run.

Original
      everyting out of date   everyting up to date
       total    user    sys    total   user    sys
--------------------------------------------------
   1  129.85   30.84   8.43    45.06   6.41   3.43
   2  109.88   29.13   9.03    45.39   6.33   3.55
   3  115.28   29.88   8.74    45.24   6.48   3.46
   4  107.67   29.41   8.29    44.31   6.03   3.55
   5  108.93   28.89   9.28    45.08   6.83   3.33
   6  107.06   29.44   8.60    44.99   5.87   3.64
   7  110.03   29.21   8.89    48.00   6.42   3.78
   8  111.24   30.61   8.80    45.87   6.33   3.57
   9  112.71   29.57   8.64    47.19   6.71   3.45
  10  110.13   29.24   8.89    44.94   6.39   3.63
  11  129.22   30.52   9.29    44.51   6.23   3.41
  12  108.10   29.93   8.57    47.04   6.33   3.83
--------------------------------------------------
avrg  113.34   29.72   8.79    45.63   6.36   3.55
sdev    7.56    0.61   0.30     1.11   0.25   0.14


Pierce
      everyting out of date   everyting up to date
       total    user    sys    total   user    sys
--------------------------------------------------
   1  110.33   30.12  19.14    48.72   9.47   6.55
   2   82.28   20.91  11.00    45.59   9.33   6.97
   3   96.81   20.93  11.74    48.37  10.57   8.43
   4   77.49   19.65   9.94    45.60   9.08   6.85
   5   88.47   22.85  10.59    48.77   9.75   7.63
   6   79.21   19.45  10.13    58.57   9.22   6.46
   7   87.43   21.04  10.75    50.50  10.32   7.08
   8   79.41   17.42   9.84    73.46  12.36   7.31
   9   87.75   22.19  11.96    50.11   9.41   7.65
  10   93.65   20.14  10.28    46.24  10.07   7.05
  11   83.44   20.68  11.02    75.97  11.96   9.40
  12   89.44   22.87  11.18    47.65   9.30   7.24
--------------------------------------------------
avrg   87.98   21.52  11.46    53.30  10.07   7.39
sdev    8.78    2.97   2.40    10.13   1.04   0.79


Mod13
      everyting out of date   everyting up to date 
       total    user    sys    total   user    sys 
--------------------------------------------------
   1   67.26    8.15   3.54    69.51   9.54   4.03
   2   62.88    8.31   3.40    77.07   9.49   4.02
   3   70.35    8.14   3.08    86.49   9.16   4.23
   4   91.94    8.80   3.34   123.81   9.21   4.34
   5   60.43    8.43   3.52    68.25   9.54   3.86
   6   57.43    8.03   3.36    63.25   9.11   4.05
   7   55.19    7.84   3.30    60.27   9.28   4.18
   8   52.44    7.71   3.33    66.78   9.45   3.85
   9   54.61    7.76   3.30    65.17   9.23   4.56
  10   65.87    7.45   3.37    64.88   9.29   3.98
  11   54.19    8.05   3.08    65.72   9.54   4.16
  12   57.69    7.90   3.33    65.87   9.13   4.01
--------------------------------------------------
avrg   62.52    8.05   3.33    73.09   9.33   4.11
sdev   10.41    0.34   0.13    16.69   0.16   0.19


Extra
      everyting out of date   everyting up to date
       total    user    sys    total   user    sys
--------------------------------------------------
   1   54.33    5.00   2.86   56.11    5.78   2.88
   2   54.48    5.42   2.59   59.92    5.92   3.22
   3   54.97    5.11   2.80   67.18    5.87   3.24
   4   55.46    4.94   2.71   56.44    5.78   3.12
   5   55.18    5.34   2.59   55.40    5.86   3.03
   6   53.98    5.39   2.61   53.92    5.60   3.29
   7   56.00    5.18   2.92   54.17    5.94   3.12
   8   51.07    5.05   2.93   53.76    5.75   3.19
   9   53.22    5.42   2.60   54.07    6.04   3.12
  10   51.33    5.31   2.91   57.47    6.01   2.96
  11   50.91    5.08   2.78   53.90    5.90   3.26
  12   52.37    5.17   2.62   54.10    5.85   3.28
--------------------------------------------------
avrg   53.61    5.20   2.74   56.37    5.86   3.14
sdev    1.72    0.16   0.13    3.72    0.12   0.13

Back to the Table of contents

Last modified: Fri Oct 31 18:58:53 EST