[gp]

I was trying to debug and improve the performance of some parallelized C++ code over the weekend for parsing CSV files. What would happen was parsing each file (~24k lines, 8 columns) would take 100ms with one execution context, but when split across many threads, the execution time of each thread would slow down proportionally and the throughput of the whole program would strictly decrease as thread count increased.

I tried all of the obvious things, but the offender ended up being a call to allocate and fill a `struct tm` object from a string representation of a date. This doesn't have any obvious reasons (to me) that it would cause cache invalidation, etc, so I'm a little in the dark.

Still, replacing this four line block improved single threaded performance by 5x, and fixed the threaded behavior, so on the whole it is now ~70x faster and parses about 400mb of csv per second.

[Thorrez]

Maybe date related code calls out to the operating system to find your time zone, and maybe that can't be done in parallel.

[Quekid5]

False sharing, maybe?

[gp]

Not a bad suggestion - thanks for the idea