Emin's email presents to me the idea of dictionaries that already contain the data we'd want to compress.  With 8 bytes of indexing data, we can refer to a TxID or a Public Key or any existing part of the blockchain.  There are also data sequences like scripts that contain a few variable chunks and are otherwise identical.  Often, the receiver has the blockchain, which contains a lot of the data that is in the message being transmitted.

First, the receiver must indicate that compressed data is preferred and the height of latest valid block it holds, and the sender must express the ability to send compressed data.  From this state, the sender sends messages that are compressed.  Compressed messages are the same as uncompressed messages except that:
  1. Data read is copied into the decompressed message until the first occurrence of 0x00, which is discarded and is followed by compressed data.
  2. Compressed data can use as a dictionary the first 16,777,215 blocks, or the last 4,244,635,647 ending with the block at the tip of the receiver's chain, or it can specify a run of zero bytes.  The sender and receiver must agree on the receiver's current block height in order to use the last 4B blocks as the dictionary.
  3. Within compressed data, the first byte identifies how to decompress:
    1. 0xFF indicates that the following three bytes are a block height with most significant byte 0x00 in network byte order.
    2. 0xFE indicates that the following byte indicates how many zero bytes to add to the decompressed data.
    3. 0xFD is an error, so compressed messages are turned off and the recipient fails the decompression process.
    4. 0x00 indicates that the zero byte by itself should be added to the decompressed data, and the data following is not compressed (return to step 1).
    5. All other values represent the most significant byte of a number to be subtracted from the receiver's current block height to identify a block height (not available until there are least 16,777,216 blocks so that this byte can be at least 0x01, since 0x00 would indicate a single zero byte, end compressed data, and return to step 1).
  4. If decompression has identified a block height (previous byte was not 0xFD, 0x00, or 0xFE), then the next four bytes identify a size (one byte) and a byte index into the block's data (three bytes), and size bytes from that block are added to the decompressed data.
  5. Steps 3 and 4 process a chunk of compressed data.  If the next byte is 0xFD, then decompression goes back to step 1 (add raw bytes until it hits a 0x00).  Otherwise, it proceeds through steps 3 (and maybe 4) again.

In Step 3.3, 0xFD causes an error, but it could be used to indicate a parameterized dictionary entry, for example 0xFD, 0x01 followed by eight more bytes to be interpreted according to steps 3.1 or 3.5 could mean OP_DUP OP_HASH160 (20 bytes from the blockchain dictionary) OP_EQUALVERIFY OP_CHECKSIG, replacing that very common occurrence of 24 bytes with 10 bytes.  Well, 11 if you include the 0x00 required by step5.  But that only works on addresses that have spent inputs.  Or 0xFD, 0x02 could be shorthand for the four zeroes of lock_time, followed by Version (1), followed by 0x01 (for one-input transactions), turning nine bytes into two for the data at the end of a normal (lock_time = 0) Txn and the beginning of a single-input Txn.  But I left 0xFD as an error because those gains didn't seem as frequent as the others.

Dave.


On Wed, Dec 2, 2015 at 3:05 PM, Peter Tschipper via bitcoin-dev <bitcoin-dev@lists.linuxfoundation.org> wrote:

On 30/11/2015 9:28 PM, Matt Corallo wrote:
I'm really not a fan of this at all. To start with, adding a compression library that is directly accessible to the network on financial software is a really, really scary idea. 
Why scary?  LZO has no current security issues, and it will be configureable by each node operator so it can be turned off completely if needed or desired. 
If there were a massive improvement, I'd find it acceptable, but the improvement you've shown really isn't all that much.
Why is 15% at the low end, to 27% at the high end not good?  It sounds like a very good boost.   
 The numbers you recently posted show it improving the very beginning of IBD somewhat over high-latency connections, but if we're throughput-limited after the very beginning of IBD, we should fix that, not compress the blocks. 
I only did the compression up to the 200,000 block to better isolate the transmission of data from the post processing of blocks and determine whether the compressing of data was adding to much to the total transmission time.

I think it's clear from the data that as the data (blocks, transactions) increase in size that (1) they compress better and (2) they have a bigger and positive impact on improving performance when compressed.

Additionally, I'd be very surprised if this had any significant effect on the speed at which new blocks traverse the network (do you have any simulations or other thoughts on this?).
From the table below, at 120000 blocks the time to sync the chain was roughly the same for compressed vs. uncompressed however after that point as block sizes start increasing, all compression libraries peformed much faster than uncompressed. The data provided in this testing clearly shows that as block size increases, the performance improvement by compressing data also increases.

TABLE 5:
Results shown in seconds with 60ms of induced latency
Num blks sync'd  Uncmp  Zlib-1  Zlib-6  LZO1x-1  LZO1x-999
---------------  -----  ------  ------  -------  ---------
120000           3226   3416    3397    3266     3302
130000           4010   3983    3773    3625     3703
140000           4914   4503    4292    4127     4287
150000           5806   4928    4719    4529     4821
160000           6674   5249    5164    4840     5314
170000           7563   5603    5669    5289     6002
180000           8477   6054    6268    5858     6638
190000           9843   7085    7278    6868     7679
200000           11338  8215    8433    8044     8795


As far as, what happens after the block is received, then obviously compression isn't going to help in post processing and validating the block, but in the pure transmission of the object it most certainly and logically does and in a fairly direct proportion to the file size (a file that is 20% smaller will be transmited "at least" 20% faster, you can use any data transfer time calculator for that).  The only issue, that I can see that required testing was to show how much compression there would be, and how much time the compression of the data would add to the sending of the data.

 

_______________________________________________
bitcoin-dev mailing list
bitcoin-dev@lists.linuxfoundation.org
https://lists.linuxfoundation.org/mailman/listinfo/bitcoin-dev




--
I like to provide some work at no charge to prove my value. Do you need a techie? 
I own Litmocracy and Meme Racing (in alpha).
I'm the webmaster for The Voluntaryist which now accepts Bitcoin.
I also code for The Dollar Vigilante.
"He ought to find it more profitable to play by the rules" - Satoshi Nakamoto