I wonder what improvements (if any) you'd see by making FastRecord a struct.

Wow, had no idea about Interlocked - thanks. I guess a tiny fraction could be saved by not parsing the 'id' to 'long'.

It looks to me that you are getting ready to hit the point where you trade readability for performance. I'm interested to see how far you are going to go with it and how far would an interviewee have to go to be considered successful.

What about using dedicate Dictionary for each thread (thus avoiding all threads contests), and then merge those dictionaries together?

reading a file will take the majority of time, parsing the strings isn't significant compared to io - why do you think running this in parallel would be the way to go? just read line by line and do the parsing immediately and you shold get more or less same speed without allocating so much memory.

Just playing with this now, feels like the class is a bit overkill. Replaced it with long[1] to see what would happen.

var stats = new ConcurrentDictionary<string, long[]>();

Parallel.ForEach(File.ReadLines(InputPath), line =>
{
    var parts = line.Split(' ');
    var entry = DateTime.Parse(parts[0]);
    var exit = DateTime.Parse(parts[1]);
    var duration = (exit - entry).Ticks;
    var id = parts[2];
    var value = stats.GetOrAdd(id, s => new[] {0L});
    value[0] += duration;
});

var summary = stats.Select(entry => $"{entry.Key} {TimeSpan.FromTicks(entry.Value[0]):c}");
File.AppendAllLines(outputPath, summary);

Took: 3,791 ms Allocated: 3,006,652 kb Peak Working Set: 39,604 kb

And with each successive iteration, as performance increases, readability decreases.

@JustPassingBy: Since you don't lock on our add (line "value[0]+=duration;") e.g. by using Interlocked.Add, the results of your code will be inaccurate.

A lot depends on your hardware as well. On my INTEL Core i5 750 2.66G 8MB BOX LGA1156 home cpu and classic HDD it produced this result: Took: 21.794 ms and allocated 2.976.442 kb with peak working set of 35.096 kb

Only after applying parallel solution, I managed to drop it down significantly. To: Took: 8.865 ms and allocated 3.030.442 kb with peak working set of 43.132 kb

@Urs: My results were consistent with previous solutions.