Okay I'm looking at the 5-chains/m and chain/d stats in debug.log. I can tweak my settings to either get higher 5-chains values OR higher chain/d values. So which should I be looking at, i.e. which is it better to have higher?
The chains/day estimate is an upcoming feature added by Sunny King. It's a probabilistic model that is supposed to give a more realistic measure of the actual chain finding performance. Right now it's not producing fully accurate estimates. On testnet the estimate was at least 12% higher than observed rate. Assuming that the model is correct, you can still use for comparative purposes including tuning the parameters. There were a few issues with the model earlier but I've been working with Sunny to fix them.
In short, chains/day should be the best performance metric so far.
That's very interesting. As I said, tweaking the sievesize/percent/roundpercent I have settings that result in higher 5-chains values but lower chains/d and vice versa, so something doesn't seem quite right there - you would expect them to both rise / fall wouldn't you?
Also why is it a decimal value (my values are in the 0.5-2.0 range)?
The problem with primes/sec and 5-chains/h is that they're not actually measuring anything useful as far as solo mining is concerned. They are more like coincidental byproducts of the mining process. Those byproducts don't really mean anything by themselves (unless you're doing pooled mining at ypool.net). You can quite easily find lots of primes/sec or 5-chains/h if you want to but solo miners should only be concerned with improving the chances to find actual blocks. That's why I performed numerous tests on testnet to test the default parameters of the client. By now it is pretty commonly known that primes/sec does not directly correlate with the block rate and I have also seen some evidence that a higher 5-chains/h does not result in more blocks.
The key idea behind the chains/day model is to estimate the probability that a candidate produces a full-length prime chain and then multiply that probability with the number of tests that are performed. As it turns out, all 3 tuning parameters supported by my client also influence this probability. They also effect the speed at which the candidates are produced and how fast they can be tested. Ideally you want to balance those settings out so that you expect to find more blocks. For example, if you adjust the settings so that the probability would be increased by 50% while the testing speed would go down by 10%, you will still get more blocks in the end. It doesn't really matter if you're finding 20% less 5-chains/h.
The reason why chains/day is a decimal number is because it's trying to represent the expected number of full-length qualifying chains that you would find in a day (i.e. chains that will result in a block if they also meet the fractional difficulty). That number is typically very low when solo mining with the current difficulty on mainnet.
? 
