Re: bohe implementation for compression tests

[Mark Nottingham <[e-mail]>]

Getting there, although you may need a small truck to haul the grain of salt that will accompany it...


On 17/01/2013, at 12:06 PM, Martin Thomson <[e-mail]> wrote:

> James,
> 
> Any feel for how bohe compares on CPU cost?  Ideally, the test harness
> would provide that information alongside the rest of the data.
> 
> On 11 January 2013 10:43, James M Snell <[e-mail]> wrote:
>> Just continuing my investigation of various header compression strategies
>> around the BOHE mechanism. In my personal github fork, I have just checked
>> in two bohe variations, one that implements selective-compression, the other
>> implements isolated-compression..
>> 
>> https://github.com/jasnell/compression-test/tree/master/compressor/bohe2
>> https://github.com/jasnell/compression-test/tree/master/compressor/bohe3
>> 
>> With bohe2 (selective-compression), a header block can consist of a
>> compressed set of headers and an uncompressed set of headers. Specific
>> headers such as Cookie, Set-Cookie, etc can be marked as "Do Not Compress".
>> These are dropped into the frame as-is and thus avoid the CRIME issue
>> completely. The rest of the headers are compressed using gzip using the
>> existing spdy3 dictionary. Obviously this is not ideal because Cookie data
>> then is passed around without any compression at all, making it far less
>> efficient than any of the other options on the table.
>> 
>> In bohe3 (isolated-compression), a header block can consist of two separate
>> compressed blocks generated using two separate stream compressor instances.
>> Selected headers (like Cookie) can be included in the secondary isolated
>> block which would never contain general user-provided header data.
>> 
>> **These are only experiments right now and are not intended as serious
>> proposals for the spec**. Isolated-compression (bohe3) does show promise
>> however. If we can successfully isolate potentially sensitive headers into
>> their own compression context generated independently of any general
>> user-supplied data, we can effectively short-circuit the CRIME attack by
>> making it impossible for an attacker to compare values based on the
>> compression ratio... and since it still uses gzip compression, we achieve a
>> generally better compression ratio overall than we get with the proposed
>> delta encoding. For now, tho, just consider this all just to be fodder for
>> discussion. There are still MANY issues with these experimental approaches
>> and I still need to go through delta in more detail to see if there is a way
>> bohe and delta can be used effectively together.
>> 
>> Just for example...
>> 
>> james-snells-macbook-pro:compression-test james$ ./compare_compressors.py -c
>> bohe3 -c bohe2 -c bohe -c delta -t
>> /Users/james/git/http_samples/mnot/amazon.com.har
>> 732 req messages processed
>>             compressed | ratio min   max   std
>> req  bohe        26,035 | 0.13  0.03  0.68  0.08
>> req bohe2        44,195 | 0.23  0.07  0.71  0.13
>> req bohe3        30,944 | 0.16  0.05  0.74  0.08
>> req delta        33,955 | 0.17  0.02  0.71  0.09
>> req http1       195,386 | 1.00  1.00  1.00  0.00
>> 
>> 732 res messages processed
>>             compressed | ratio min   max   std
>> res  bohe        39,525 | 0.25  0.04  0.67  0.07
>> res bohe2        47,157 | 0.29  0.12  0.71  0.08
>> res bohe3        44,843 | 0.28  0.06  0.70  0.07
>> res delta        44,499 | 0.28  0.02  0.65  0.09
>> res http1       159,968 | 1.00  1.00  1.00  0.00
>> 
>> - James
>> 
>> 
>> 
>> On Thu, Jan 10, 2013 at 11:08 AM, James M Snell <[e-mail]> wrote:
>>> 
>>> I have an initial bohe implementation for the compression tests... it's
>>> very preliminary and uses the same gzip compression as the current spdy3.
>>> I'm going to be playing around with the delta compression mechanism as well
>>> and see how much of an impact that has. Initial results are very promising
>>> but I haven't done much debugging yet. Just wanted folks to know that this
>>> work was underway...
>>> 
>>> https://github.com/jasnell/compression-test/tree/master/compressor/bohe
>>> 
>>> Some test runs....
>>> 
>>> ./compare_compressors.py -c bohe -c spdy3 -c delta
>>> ../http_samples/mnot/amazon.com.har
>>> 732 req messages processed
>>>             compressed | ratio min   max   std
>>> req  bohe        26,122 | 0.13  0.04  0.70  0.08
>>> req delta        33,955 | 0.17  0.02  0.71  0.09
>>> req http1       195,386 | 1.00  1.00  1.00  0.00
>>> req spdy3        27,238 | 0.14  0.04  0.71  0.08
>>> 
>>> 732 res messages processed
>>>             compressed | ratio min   max   std
>>> res  bohe        39,628 | 0.25  0.04  0.66  0.07
>>> res delta        44,499 | 0.28  0.02  0.65  0.09
>>> res http1       159,968 | 1.00  1.00  1.00  0.00
>>> res spdy3        41,325 | 0.26  0.04  0.67  0.08
>>> 
>>> 
>>> ./compare_compressors.py -c bohe -c spdy3 -c delta
>>> ../http_samples/mnot/craigslist.org.har
>>> 66 req messages processed
>>>             compressed | ratio min   max   std
>>> req  bohe         1,948 | 0.15  0.06  0.73  0.11
>>> req delta         2,036 | 0.16  0.07  0.71  0.11
>>> req http1        12,894 | 1.00  1.00  1.00  0.00
>>> req spdy3         2,016 | 0.16  0.07  0.75  0.11
>>> 
>>> 66 res messages processed
>>>             compressed | ratio min   max   std
>>> res  bohe         1,786 | 0.18  0.07  0.77  0.13
>>> res delta         2,858 | 0.28  0.08  0.69  0.12
>>> res http1        10,147 | 1.00  1.00  1.00  0.00
>>> res spdy3         1,869 | 0.18  0.09  0.78  0.13
>>> 
>>> 
>>> ./compare_compressors.py -c bohe -c spdy3 -c delta
>>> ../http_samples/mnot/flickr.com.har
>>> 438 req messages processed
>>>             compressed | ratio min   max   std
>>> req  bohe        11,988 | 0.10  0.02  0.69  0.07
>>> req delta        26,372 | 0.22  0.01  0.71  0.14
>>> req http1       121,854 | 1.00  1.00  1.00  0.00
>>> req spdy3        12,550 | 0.10  0.02  0.71  0.07
>>> 
>>> 438 res messages processed
>>>             compressed | ratio min   max   std
>>> res  bohe        13,073 | 0.09  0.05  0.66  0.06
>>> res delta        25,236 | 0.18  0.02  0.70  0.11
>>> res http1       140,457 | 1.00  1.00  1.00  0.00
>>> res spdy3        14,142 | 0.10  0.05  0.66  0.06
>>> 
>>> 
>>> ./compare_compressors.py -c bohe -c spdy3 -c delta
>>> ../http_samples/mnot/facebook.com.har
>>> 234 req messages processed
>>>             compressed | ratio min   max   std
>>> req  bohe         6,091 | 0.15  0.06  0.78  0.07
>>> req delta         7,800 | 0.19  0.02  0.70  0.07
>>> req http1        41,980 | 1.00  1.00  1.00  0.00
>>> req spdy3         6,301 | 0.15  0.06  0.77  0.07
>>> 
>>> 234 res messages processed
>>>             compressed | ratio min   max   std
>>> res  bohe         9,458 | 0.23  0.07  0.68  0.07
>>> res delta        12,045 | 0.30  0.13  0.60  0.08
>>> res http1        40,252 | 1.00  1.00  1.00  0.00
>>> res spdy3         9,788 | 0.24  0.07  0.69  0.07
>>> 
>>> 
>>> 
>>> 
>> 
> 

--
Mark Nottingham   http://www.mnot.net/