Martin Kersten wrote:
Consequently, moving to a binary protocol is not warranted by comparison, nor effective in terms of performance improvement. 50% of the cost in MonetDB for these queries are the SQL parser and query cache matching.
Given the focus of MonetDB on datawarehousing, rather then high volume web-interactions, there has been no steps taken to improve the protocol.
You only mention here the performance gain within MonetDB, as opposed to the performance gain that could be generated if the client doesn't need to parse integers from strings to work on them but only make an in memory reference to the current row, and threating it as a c-struct. Indeed. Even in an application it becomes relevant to know how much time is actually spent at the interface. My conjuncture is
Stefan de Konink wrote: that it is a few percent only, contrived cases excluded. Fastly outnumbered by even the network delays. Measuring applications and studying there behavior using e.g. Callgrind gives you a reasonable insight in the performance bottlenecks. Eg. in your recent callgrind file there was about a 10% cpu cycles loss due to virtualization it seemed.
Stefan