aboutsummaryrefslogtreecommitdiff
path: root/doc/parallelism.txt
blob: e509d8631ad1e883475d0311737ba0cc3864b906 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163

                      Parallelizing SquashFS Data Packing
                      ***********************************

 0) Overview
 ***********

 On a high level, data blocks are processed as follows:

 The "block processor" has a simple begin/append/end interface for submitting
 file data. Internally it chops the file data up into fixed size blocks that
 are each [optionally] compressed and hashed. If the "end" function is called
 and there is still left over data, a fragment is created.

 Fragments are only hashed. If another fragment exists with the same size and
 hash, it is discarded and the existing fragment is referenced. Fragments are
 collected in a fragment block that, once it overflows, is processed like a
 normal block.

 The final compressed & hashed data blocks & fragment blocks are passed on to
 the "block writer".

 The block writer simply writes blocks to the output file. Flags are used to
 communicate what the first and last block of a file are. Entire files are
 deduplicated by trying to find a sequence of identical size/hash pairs in
 the already written blocks.


 0.1) Implementation

 The implementation of the block processor is in lib/sqfs/block_processor. The
 file common.c contains the frontend for file data submission and common
 functions for processing a single block, handling a completed block and
 handling a completed fragment.

 A reference serial implementation is provided in the file serial.c


 1) Thread Pool Based Block Processor
 ************************************

 The main challenge of parallelizing the block processor lies in the fact the
 output HAS TO BE byte-for-byte equivalent to the serial reference
 implementation.

 This means:
  - File blocks have to be written in the exact same order as they
    are submitted.
  - If storing a fragment overflows the fragment block, the resulting
    fragment block has to be written next, no file data.


 The current implementation in winpthread.c (based on pthread or Windows
 native threads, depending on whats available) uses the following approach:

  - Each submitted data block or fragment gets an incremental sequence number
    and is appended to a FIFO queue.
  - Multiple threads consume blocks from the queue and use the function
    from common.c to process the dequeued blocks.
  - Completed blocks are inserted into a "done" queue, sorted by their
    sequence number.
  - The main thread that submits blocks also dequeues the completed ones,
    keeping track of the sequence numbers, and calls the respective common
    functions for processing completed blocks and fragments.
  - If a fragment block is created, it is submitted with *the same* sequence
    number as the fragment that caused it to overflow and the next expected
    sequence number is reset to that.

 To make sure the queue doesn't fill all RAM, submitted blocks are counted.
 The counter is decremented when dequeueing completed blocks. If it reaches
 a maximum, signal/await is used to wait for the worker threads to complete
 some blocks to process. Similarly, the worker threads use signal/await to
 wait on the queue if it is empty.


 1.1) Problems

 The outlined approach performs sub-optimal, with an efficiency somewhere
 between 50% to 75% on the bench mark data used.

 Profiling using perf shows that almost a third of the time, only one
 worker thread is actually active, while the others are waiting.

 The current hypothesis is that this is caused by many small input files
 being processed, causing a work load consisting primarily of fragments.

  - Fragments are only hashed, not compressed, so the work is
    primarily I/O bound.
  - After a number of fragments are consumed, a fragment block is created.
  - The fragment block is submitted to the almost empty queue and the
    I/O thread waits for it to be completed before doing anything else.
  - One thread gets to handle the fragment block, which involves a lot more
    work. Meanwhile the other threads starve on the empty queue.
  - After that has finally been handed of to the I/O thread, another burst of
    fragments comes in.
  - Rinse and repeat.


 1.2) Proposed Solution

 It makes no sense for the main thread to block until the fragment block is
 done. It can process further fragments (just not write blocks), creating
 more fragment blocks on the way.

 A possible implementation might be to maintain 3 queues instead of 2:

  - A queue with submitted blocks.
  - A queue with completed blocks.
  - A queue for blocks ready to be written to disk ("I/O queue").

 A second sequence number is needed for keeping order on the I/O queue:

  - Submit blocks as before with incremental processing sequence number.
  - Dequeue completed blocks in order by processing sequence number.
     - For regular blocks, add them to the I/O queue with incremental I/O
       sequence number.
     - For fragments, consolidate them into fragment blocks. On overflow,
       dispatch a fragment block with incremental processing sequence number,
       BUT give it an I/O queue sequence number NOW.
     - For fragment blocks, add them to the I/O queue without allocating an
       I/O sequence number, it already has one.
  - Dequeue ordered by I/O sequence number from the I/O queue and send the
    completed blocks to the block writer.


 If you have a more insights or a better idea, please let me know.



 2) Benchmarks
 *************

 TODO: benchmarks with the following images:
        - Debian live iso (2G)
	- Arch Linux live iso (550M)
	- Raspberry Pi 3 QT demo image (~300M)

       sqfs2tar $IMAGE | tar2sqfs -j $NUM_CPU -f out.sqfs

       Values to measure:
        - Total wall clock time of tar2sqfs.
	- Througput (bytes read / time, bytes written / time).

       Try the above for different compressors and stuff everything into
       a huge spread sheet. Then, determine the following and plot some
       nice graphs:

        - Absolute speedup (normalized to serial implementation).
	- Absolute efficiency (= speedup / $NUM_CPU)
        - Relative speedup (normalized to thread pool with -j 1).
	- Relative efficiency


 Available test hardware:
  - 8(16) core AMD Ryzen 7 3700X, 32GiB DDR4 RAM.
  - Various 4 core Intel Xeon servers. Precise Specs not known yet.
  - TODO: Check if my credentials on LCC2 still work. The cluster nodes AFAIK
    have dual socket Xeons. Not sure if 8 cores per CPU or 8 in total?

 For some compressors and work load, tar2sqfs may be I/O bound rather than CPU
 bound. The different machines have different storage which may impact the
 result. Should this be taken into account for comparison or eliminated by
 using a ramdisk or fiddling with the queue backlog?