A.1 Reply and Object Sizes

only for RuBoard - do not distribute or recompile

A.1 Reply and Object Sizes

Figure A-1 and Figure A-2 show two size distributions, the first for objects and the second for reply sizes. The reply size distribution includes all responses sent to clients , and repeated URLs are counted each time. The object size distribution is meant to represent the size of objects stored on origin servers. Thus, repeated URLs are filtered out for the object size data so that each URL is counted just once. The reply size distribution has a mean of 8,419 bytes, and the median is 1,266 bytes. The object size distribution is quite different, with a mean of 22,499 bytes and a median of 3,372 bytes.

Figure A-1. Distribution of object sizes (IRCache data)

Figure A-2. Distribution of reply sizes (IRCache data)

Both distributions are derived from the fifth field of Squid's access.log file. For reply sizes, I include almost every response. I filter out TCP_DENIED responses and certain other Squid-specific responses (e.g., Cache Digests). After that, I take all the size values and calculate the mean, median, and a histogram. Calculating the object size distribution is a little trickier. First, I filter out non-GET requests and non-200 status responses. This is cheating a little bit because POST requests , for example, have origin server objects, or resources, that aren't included in this analysis. Second, I filter out repeated URLs. For each URL, I keep the largest size value seen in the access log. Finally, these values are used to calculate the mean, median, and histogram.

As you can see, reply sizes tend to be smaller than object sizes. This is because a significant percentage of the responses in the reply size distribution don't have an HTTP message body ”for example, 304 (Not Modified) responses and 404 (Not Found) error messages. As we'll see later, about 20% of all responses in the IRCache traces are 304 messages.

only for RuBoard - do not distribute or recompile