3 ms·
Duplicate. See lively discussion at http://news.ycombinator.com/item?id=1570694 http://news.ycombinator.com/item?id=1570694
by wglb 16y ago
Duplicate.
See lively discussion at http://news.ycombinator.com/item?id=1570694 http://news.ycombinator.com/item?id=1570694
- jacquesm 16y agoNot exactly, this is where Zed presumably shows that using pipes is equivalent to real world testing and has nothing to do with using the loopback/localhost device. Personally I think that's nonsense. I downloaded the code, checked it out, there is nothing in there that changes my mind on the whole thing, the meat is in the C code and it's the same as what was linked from the article. The rest is a wrapper script and some stuff to visualise the results, which are what he showed them to be. The premise that web servers have the majority of their fds active is where the issue lies. FWIW I did some measuring and the results are pretty much in line with what I expected (though the spread is still puzzling, I really intend to find out the cause of that): http://news.ycombinator.com/item?id=1573145 http://news.ycombinator.com/item?id=1573145
- lol 16y agoThe massive thing that you're ignoring is that he's trying to make mongrel2 useful as more than a standard web server. Sure, if you ignore that massively obvious piece of information, you have a point. But if you take it into account, you really don't have one at all. As so often happens, you've spewed a lot of words, but said absolutely nothing of value.
- zedshaw 16y agoIf the test harness is wrong than every other test that everyone doing epoll comparisons with poll has done is wrong. This is the one they've all been using. So, either you write your own test to evaluate the same thing, without confounding it with other external factors (like this test does), or you use this test and prove me wrong by finding different data. However, I'm glad you finally went and got some metrics to show one webserver instead of the useless rhetoric you were spewing before. Not that it proves a damn thing since it has no external validity, but at least it's a step in the right direction. I also say that your findings of 40% are much higher than what you implied in your previous comments. You made it seem like it's close to the 10% range, but 40% is damn near ready to leap over into the equal performance levels. It's interesting that even after that you still can't see the value of either making epoll faster or using both, but whatever, you can't convince every idiot on the internet, even with Science.
- jacquesm 16y agoYou should read more carefully. 40% was the highest observed, 10% the lowest. Also, I said 'the majority', which is not like making it seem like it is only 10%. So it's not '10%' or '40%', for this particular workload, on average you're hovering somewhere between 25 and 30%, since you seem to prefer the higher number make that 30%. And if you read a bit more you would see that there are some factors that could pull that number down considerably in different situations and only one major one (large file serving, streaming) that would pull it up. If I get around to it (I'm on a holiday right now) I'm going to instrument some more sites to measure their active-to-total ratios to see if there is some kind of number (or a set of them) that we might be able to agree on when it comes to real life web traffic. One of them is a filedump, which serves up large media files to clients, that's the use case that I can come up with in which your scheme might pay off, and in which it is possibly advantageous (even if it is a very small advantage) to use poll. You keep 'challenging' people to come up with numbers. So, I did, from a real life webserver serving real customers, lots of them. If 60% is the cutoff point (and we're assuming that you are right about that) then 40% is comfortably under that, not 'much higher' than what I wrote in my previous comments, the majority (60 to 90%) of the fds is idle. So it's not in to 'equal performance levels' yet. But as far as I'm concerned even that is probably moot. (But that needs more testing) because I suspect that even this fairly high traffic server spends only very small amount of time on polling the OS, and a much larger time on shovelling bits in and out of the system. You keep using that word 'science' as though measuring anything at all is automatically science. But what you're doing is more like selecting the data that fits your 'who needs epoll' and 'epoll is bad' hypothesis and rejecting out of hand any experience and / or experimental data that seems to prove you wrong. That's not science. Science is to use all the data available to you, not just the juicy bits. Not just those bits that allow you to make bold claims about discoveries when in fact all you've done is to show that the underlying implementation of one systemcall is different and has different trade-offs than another, which may be specific to the way your system is set up. I agree that to test something you need to reduce it to the minimum of code required to do your tests. But when you're going that route you are doing what so many other database, CPU and application benchmarkers forget. The real world is nothing like a benchmark. You to suggest that this is 'not a localhost test' (http://news.ycombinator.com/item?id=1573195 http://news.ycombinator.com/item?id=1573195) and you say so pretty loud. That would mean that there may be some validity to your data that I have not been able to find to date in the stuff you've provided. I'm challenging you to provide the data that apparently is still missing that proves that this was not 'a localhost test'. And if you don't supply that data then I think you should take back (and try to do it gracriously for a change) that this wasn't a local host test, or you're going to have to come with some definition of what you consider a localhost test to be. For me, that definition is simple, a localhost test is a benchmark that never generates any network traffic. Feel free to disagree.
- lrm242 16y agoNot a dupe -- this is his follow-up post.
- wglb 16y agoMy bad. This is a follow-up article. Apologies.