Re: Interesting post about AIO
Hi, Agreeing with jeff, you can't complain of the lack of an accessor, when you just voided the attachment ;) I really wonder what JDK guys have in mind with AIO, it's far of being a slick API, it doesn't support UDP (you need to mix AIO and NIO if you want to support TCP and UDP) and I can't see how it could be a performance gain ? After few year of usage I start to think the whole java.nio and java.concurency are weirdly designed API. Anyway we can't expect more performance for AIO since it's probably just an API built on top of NIO. If we want more performance, we need to zero copy interface with the linux epoll() sys call. But I'm not sure we want to do that :) The funny point is the message of the guy asking does jetty scale like vert.x :) IT peoples are pretty stupid sometimes and trust any crappy benchmark buzz. Julien On Fri, Mar 1, 2013 at 5:45 PM, Emmanuel Lécharny elecha...@gmail.comwrote: Le 3/1/13 5:00 PM, Jeff MAURY a écrit : I don't quite follow his arguments because this is easily fixed by either using an anonymous class or a dedicated attachment. I'm more concern about the need to allocate a buffer for every read (a clear no go from my POV) an about the write handling. I would like to see a chapter on performance as my POC is not really good on this point and I would like to know if the cause is clearly from NIO or from my MINA adapter on top on NIO More to come... I guess the only way to know about that is first to build a prototype to compare it with what we have and probably by looking at AIO code to see if it's just a wrapper on top of NIO or something smarter that uses the underlying OS asunc IO... -- Regards, Cordialement, Emmanuel Lécharny www.iktek.com
Fwd: Re: Interesting post about AIO
forwarding to the correct ML -- Forwarded message -- From: Emmanuel Lécharny elecha...@gmail.com Date: Mar 2, 2013 2:13 PM Subject: Re: Interesting post about AIO To: Apache Directory Developers List d...@directory.apache.org Le 3/2/13 9:10 AM, Julien Vermillard a écrit : Hi, Agreeing with jeff, you can't complain of the lack of an accessor, when you just voided the attachment ;) Again, this is not the interesting part n this article... I really wonder what JDK guys have in mind with AIO, it's far of being a slick API, it doesn't support UDP (you need to mix AIO and NIO if you want to support TCP and UDP) and I can't see how it could be a performance gain ? Java AIO implementation is supposed to benefit from OS taht has a specific way to deal with AIO (through completion queues, like what is done on Windows, some Linux versions and probably other OS). This page is providing some information : http://www.kegel.com/c10k.html#aio Another very complete description : http://stackoverflow.com/questions/4093185/whats-the-difference-between-epoll-poll-threadpool Regarding AIO/ePoll on linux, an interesting paper : https://docs.google.com/viewer?a=vq=cache:CwPj9geCCTkJ:www.linuxsymposium.org/archives/OLS/Reprints-2002/lahaise-reprint.pdf+hl=engl=frpid=blsrcid=ADGEESj6bzG_kAkqH8KQGWL5Evt4yDbIEhrsaaNMD4eD7p-bhJHoAjjQeT3Ys8fXWY_dn88WY98kW0Hg2-sBIPrvFpBpniMrKK6wY07lY80x1ZT9tEX_TERnxwosL6WXjokFs_huSdX6sig=AHIEtbRAgWiO-HR0BOjkxh3JcYetJpTXnA After few year of usage I start to think the whole java.nio and java.concurency are weirdly designed API. That's probably why API like MINA are necessary :) Anyway we can't expect more performance for AIO since it's probably just an API built on top of NIO. Not so sure. If we want more performance, we need to zero copy interface with the linux epoll() sys call. But I'm not sure we want to do that :) AIO does no copy the data, compared to NIO. It's probably one of the aspect that leads to better performances. I'm still wondering if we can see any gain, but the more I read about AIO, the more I'm convinced we need to spend some time on a prototype to get valid numbers. In any case, would we implement a layer on top of AIO, the upper API should not be different, and that's were we have to be careful ! -- Regards, Cordialement, Emmanuel Lécharny www.iktek.com
Re: Interesting post about AIO
On Sat, Mar 2, 2013 at 9:43 AM, Emmanuel Lécharny elecha...@gmail.comwrote: Le 3/2/13 9:10 AM, Julien Vermillard a écrit : Hi, Agreeing with jeff, you can't complain of the lack of an accessor, when you just voided the attachment ;) Again, this is not the interesting part n this article... I really wonder what JDK guys have in mind with AIO, it's far of being a slick API, it doesn't support UDP (you need to mix AIO and NIO if you want to support TCP and UDP) and I can't see how it could be a performance gain ? Java AIO implementation is supposed to benefit from OS taht has a specific way to deal with AIO (through completion queues, like what is done on Windows, some Linux versions and probably other OS). This is not yet done in the current implementation of the JDK. Maybe JDK8 ? This page is providing some information : http://www.kegel.com/c10k.html#aio Another very complete description : http://stackoverflow.com/questions/4093185/whats-the-difference-between-epoll-poll-threadpool Regarding AIO/ePoll on linux, an interesting paper : https://docs.google.com/viewer?a=vq=cache:CwPj9geCCTkJ:www.linuxsymposium.org/archives/OLS/Reprints-2002/lahaise-reprint.pdf+hl=engl=frpid=blsrcid=ADGEESj6bzG_kAkqH8KQGWL5Evt4yDbIEhrsaaNMD4eD7p-bhJHoAjjQeT3Ys8fXWY_dn88WY98kW0Hg2-sBIPrvFpBpniMrKK6wY07lY80x1ZT9tEX_TERnxwosL6WXjokFs_huSdX6sig=AHIEtbRAgWiO-HR0BOjkxh3JcYetJpTXnA After few year of usage I start to think the whole java.nio and java.concurency are weirdly designed API. That's probably why API like MINA are necessary :) Anyway we can't expect more performance for AIO since it's probably just an API built on top of NIO. Not so sure. It's not done on top of NIO but the current implementation (JDK7) does not use OS based NIO but does more or less what MINA NIO does If we want more performance, we need to zero copy interface with the linux epoll() sys call. But I'm not sure we want to do that :) AIO does no copy the data, compared to NIO. It's probably one of the aspect that leads to better performances. I don't see the difference between NIO and AIO on this topic as both API use the ByteBuffer for exchanging data. I'm still wondering if we can see any gain, but the more I read about AIO, the more I'm convinced we need to spend some time on a prototype to get valid numbers. I don't expect good numbers given the current implementation but at least we've got something when the JDK implementation will be better. In any case, would we implement a layer on top of AIO, the upper API should not be different, and that's were we have to be careful ! Yes, that what I have in my POC Jeff -- Regards, Cordialement, Emmanuel Lécharny www.iktek.com -- Jeff MAURY Legacy code often differs from its suggested alternative by actually working and scaling. - Bjarne Stroustrup http://www.jeffmaury.com http://riadiscuss.jeffmaury.com http://www.twitter.com/jeffmaury
Re: Interesting post about AIO
Le 3/2/13 1:23 PM, Jeff MAURY a écrit : On Sat, Mar 2, 2013 at 9:43 AM, Emmanuel Lécharny elecha...@gmail.comwrote: Le 3/2/13 9:10 AM, Julien Vermillard a écrit : Hi, Agreeing with jeff, you can't complain of the lack of an accessor, when you just voided the attachment ;) Again, this is not the interesting part n this article... I really wonder what JDK guys have in mind with AIO, it's far of being a slick API, it doesn't support UDP (you need to mix AIO and NIO if you want to support TCP and UDP) and I can't see how it could be a performance gain ? Java AIO implementation is supposed to benefit from OS taht has a specific way to deal with AIO (through completion queues, like what is done on Windows, some Linux versions and probably other OS). This is not yet done in the current implementation of the JDK. Maybe JDK8 ? We have to check if JDK8 is any different. This page is providing some information : http://www.kegel.com/c10k.html#aio Another very complete description : http://stackoverflow.com/questions/4093185/whats-the-difference-between-epoll-poll-threadpool Regarding AIO/ePoll on linux, an interesting paper : https://docs.google.com/viewer?a=vq=cache:CwPj9geCCTkJ:www.linuxsymposium.org/archives/OLS/Reprints-2002/lahaise-reprint.pdf+hl=engl=frpid=blsrcid=ADGEESj6bzG_kAkqH8KQGWL5Evt4yDbIEhrsaaNMD4eD7p-bhJHoAjjQeT3Ys8fXWY_dn88WY98kW0Hg2-sBIPrvFpBpniMrKK6wY07lY80x1ZT9tEX_TERnxwosL6WXjokFs_huSdX6sig=AHIEtbRAgWiO-HR0BOjkxh3JcYetJpTXnA After few year of usage I start to think the whole java.nio and java.concurency are weirdly designed API. That's probably why API like MINA are necessary :) Anyway we can't expect more performance for AIO since it's probably just an API built on top of NIO. Not so sure. It's not done on top of NIO but the current implementation (JDK7) does not use OS based NIO but does more or less what MINA NIO does That's not exactly true. If you look at OpenJDK 7 code, at least on Windows, AIO uses CompletionPort (the WindowsAsynchronousChannelProvider class depends on Iocp class). On MacOSX, I'm not sure if it's not using a wrapper on top of NIO (seems likely) We can ignore Microsoft implementation, but MO, that would be a mistake... -- Regards, Cordialement, Emmanuel Lécharny www.iktek.com
Re: Fwd: Re: Interesting post about AIO
Thanks for the pointers Kiran I m going to read it. Julien Le 2 mars 2013 12:46, Kiran Ayyagari kayyag...@apache.org a écrit : forwarding to the correct ML -- Forwarded message -- From: Emmanuel Lécharny elecha...@gmail.com Date: Mar 2, 2013 2:13 PM Subject: Re: Interesting post about AIO To: Apache Directory Developers List d...@directory.apache.org Le 3/2/13 9:10 AM, Julien Vermillard a écrit : Hi, Agreeing with jeff, you can't complain of the lack of an accessor, when you just voided the attachment ;) Again, this is not the interesting part n this article... I really wonder what JDK guys have in mind with AIO, it's far of being a slick API, it doesn't support UDP (you need to mix AIO and NIO if you want to support TCP and UDP) and I can't see how it could be a performance gain ? Java AIO implementation is supposed to benefit from OS taht has a specific way to deal with AIO (through completion queues, like what is done on Windows, some Linux versions and probably other OS). This page is providing some information : http://www.kegel.com/c10k.html#aio Another very complete description : http://stackoverflow.com/questions/4093185/whats-the-difference-between-epoll-poll-threadpool Regarding AIO/ePoll on linux, an interesting paper : https://docs.google.com/viewer?a=vq=cache:CwPj9geCCTkJ:www.linuxsymposium.org/archives/OLS/Reprints-2002/lahaise-reprint.pdf+hl=engl=frpid=blsrcid=ADGEESj6bzG_kAkqH8KQGWL5Evt4yDbIEhrsaaNMD4eD7p-bhJHoAjjQeT3Ys8fXWY_dn88WY98kW0Hg2-sBIPrvFpBpniMrKK6wY07lY80x1ZT9tEX_TERnxwosL6WXjokFs_huSdX6sig=AHIEtbRAgWiO-HR0BOjkxh3JcYetJpTXnA After few year of usage I start to think the whole java.nio and java.concurency are weirdly designed API. That's probably why API like MINA are necessary :) Anyway we can't expect more performance for AIO since it's probably just an API built on top of NIO. Not so sure. If we want more performance, we need to zero copy interface with the linux epoll() sys call. But I'm not sure we want to do that :) AIO does no copy the data, compared to NIO. It's probably one of the aspect that leads to better performances. I'm still wondering if we can see any gain, but the more I read about AIO, the more I'm convinced we need to spend some time on a prototype to get valid numbers. In any case, would we implement a layer on top of AIO, the upper API should not be different, and that's were we have to be careful ! -- Regards, Cordialement, Emmanuel Lécharny www.iktek.com
Re: Fwd: Re: Interesting post about AIO
Maybe this is an interesting read about AIO too. Its written by one of the jetty developers. http://webtide.intalio.com/2013/02/on-jdk-7-asynchronous-io/ On Sat, Mar 2, 2013 at 5:42 PM, Julien Vermillard jvermill...@gmail.comwrote: Thanks for the pointers Kiran I m going to read it. Julien Le 2 mars 2013 12:46, Kiran Ayyagari kayyag...@apache.org a écrit : forwarding to the correct ML -- Forwarded message -- From: Emmanuel Lécharny elecha...@gmail.com Date: Mar 2, 2013 2:13 PM Subject: Re: Interesting post about AIO To: Apache Directory Developers List d...@directory.apache.org Le 3/2/13 9:10 AM, Julien Vermillard a écrit : Hi, Agreeing with jeff, you can't complain of the lack of an accessor, when you just voided the attachment ;) Again, this is not the interesting part n this article... I really wonder what JDK guys have in mind with AIO, it's far of being a slick API, it doesn't support UDP (you need to mix AIO and NIO if you want to support TCP and UDP) and I can't see how it could be a performance gain ? Java AIO implementation is supposed to benefit from OS taht has a specific way to deal with AIO (through completion queues, like what is done on Windows, some Linux versions and probably other OS). This page is providing some information : http://www.kegel.com/c10k.html#aio Another very complete description : http://stackoverflow.com/questions/4093185/whats-the-difference-between-epoll-poll-threadpool Regarding AIO/ePoll on linux, an interesting paper : https://docs.google.com/viewer?a=vq=cache:CwPj9geCCTkJ:www.linuxsymposium.org/archives/OLS/Reprints-2002/lahaise-reprint.pdf+hl=engl=frpid=blsrcid=ADGEESj6bzG_kAkqH8KQGWL5Evt4yDbIEhrsaaNMD4eD7p-bhJHoAjjQeT3Ys8fXWY_dn88WY98kW0Hg2-sBIPrvFpBpniMrKK6wY07lY80x1ZT9tEX_TERnxwosL6WXjokFs_huSdX6sig=AHIEtbRAgWiO-HR0BOjkxh3JcYetJpTXnA After few year of usage I start to think the whole java.nio and java.concurency are weirdly designed API. That's probably why API like MINA are necessary :) Anyway we can't expect more performance for AIO since it's probably just an API built on top of NIO. Not so sure. If we want more performance, we need to zero copy interface with the linux epoll() sys call. But I'm not sure we want to do that :) AIO does no copy the data, compared to NIO. It's probably one of the aspect that leads to better performances. I'm still wondering if we can see any gain, but the more I read about AIO, the more I'm convinced we need to spend some time on a prototype to get valid numbers. In any case, would we implement a layer on top of AIO, the upper API should not be different, and that's were we have to be careful ! -- Regards, Cordialement, Emmanuel Lécharny www.iktek.com
Re: Fwd: Re: Interesting post about AIO
Le 3/2/13 7:01 PM, Christian Schwarz a écrit : Maybe this is an interesting read about AIO too. Its written by one of the jetty developers. http://webtide.intalio.com/2013/02/on-jdk-7-asynchronous-io/ Hmmm, this is the exact link that started this discussion :-) -- Regards, Cordialement, Emmanuel Lécharny www.iktek.com
