zhangxiang
2008-May-26 01:38 UTC
[Xen-devel] Why the credit scheduler of xen-3.2.0 can ensure the I/O fairness?
Hello, everybody: I have recently read the paper "Scheduling I/O in Virtual Machine Monitors", which said that the credit scheduler of xen can not ensure the i/o fairness(Please see the Figure 3 in the paper). The author said that you can achieve the i/o fairness by disable the scheduler tickling, and sorting the vcpu by the value of the credit remained(Please see the relevant ppt for VEE08). I have evaluated the credit scheduler of xen-3.1.0, it does agree with the option in the paper. However, when I evaluating the credit scheduler of xen-3.2.0, the results show that the scheduler can ensure the i/o fairness very well. I diff the sched_credit.c of the two versions of xen, they are the same. It seems that the credit scheduler of xen-3.2.0 does not sort the vcpu by credit or disable the scheduler tickling. So I wonder that what improvements xen have done to solove this problem? The experiment I have done is as follows: The system has an AMD Opteron 848 processor, 4G physical memory, and a Gigabit ethernet NIC. I run eight xen VMs, and run netperf benchmark separatly. Every VM has almost the same bandwidth of 120 Mb/s. Thank you very much for your help. Xiang ZHANG National Research Center for Intelligent Computing Systems Institute of Computing Technology, Chinese Academy of Sciences, Beijing, P.R.China _______________________________________________ Xen-devel mailing list Xen-devel@lists.xensource.com http://lists.xensource.com/xen-devel
Atsushi SAKAI
2008-May-26 02:05 UTC
Re: [Xen-devel] Why the credit scheduler of xen-3.2.0 can ensure theI/O fairness?
Hi, Xiang You should see the following patch. changeset: Process event channels notifications in round-robin order. changeset 324: 7fe1c6d02a2b http://xenbits.xensource.com/linux-2.6.18-xen.hg?rev/7fe1c6d02a2b Thanks Atsushi SAKAI "zhangxiang" <zhangxiang@ncic.ac.cn> wrote:> Hello, everybody: > > I have recently read the paper "Scheduling I/O in Virtual Machine Monitors", which said that the credit scheduler of xen can not ensure the i/o fairness(Please see the Figure 3 in the paper). The author said that you can achieve the i/o fairness by disable the scheduler tickling, and sorting the vcpu by the value of the credit remained(Please see the relevant ppt for VEE08). I have evaluated the credit scheduler of xen-3.1.0, it does agree with the option in the paper. > > However, when I evaluating the credit scheduler of xen-3.2.0, the results show that the scheduler can ensure the i/o fairness very well. I diff the sched_credit.c of the two versions of xen, they are the same. It seems that the credit scheduler of xen-3.2.0 does not sort the vcpu by credit or disable the scheduler tickling. So I wonder that what improvements xen have done to solove this problem? > > The experiment I have done is as follows: > The system has an AMD Opteron 848 processor, 4G physical memory, and a Gigabit ethernet NIC. I run eight xen VMs, and run netperf benchmark separatly. Every VM has almost the same bandwidth of 120 Mb/s. > > Thank you very much for your help. > > Xiang ZHANG > National Research Center for Intelligent Computing Systems > Institute of Computing Technology, Chinese Academy of Sciences, Beijing, > P.R.China_______________________________________________ Xen-devel mailing list Xen-devel@lists.xensource.com http://lists.xensource.com/xen-devel
zhangxiang
2008-May-26 02:31 UTC
Re: Re: [Xen-devel] Why the credit scheduler of xen-3.2.0 can ensure theI/O fairness?
Hello Atsushi: Thank you very much for your reply! I have already read that patch. In the paper "Scheduling I/O in Virtual Machine Monitors", the author said that fixing event channel notification can only ensure the I/O fairness with the tickling scheduler disable, and the scheduler tickling is the culprit of the I/O unfairness even the event channel code is fixed (Please see the Section 6.1). Although Xen-3.2.0 have fixxed the event channel code, it does not disable the scheduler tickling. So I think there should be something else xen-3.2.0 has done. Thank you again for your help. Xiang ZHANG National Research Center for Intelligent Computing Systems Institute of Computing Technology, Chinese Academy of Sciences, Beijing, P.R.China 发件人： Atsushi SAKAI 发送时间： 2008-05-26 09:56:35 收件人： zhangxiang 抄送： xen-devel 主题： Re: [Xen-devel] Why the credit scheduler of xen-3.2.0 can ensure theI/O fairness? Hi, Xiang You should see the following patch. changeset: Process event channels notifications in round-robin order. changeset 324: 7fe1c6d02a2b http://xenbits.xensource.com/linux-2.6.18-xen.hg?rev/7fe1c6d02a2b Thanks Atsushi SAKAI "zhangxiang" <zhangxiang@ncic.ac.cn > wrote:> Hello, everybody: > > I have recently read the paper "Scheduling I/O in Virtual Machine Monitors", which said that the credit scheduler of xen can not ensure the i/o fairness(Please see the Figure 3 in the paper). The author said that you can achieve the i/o fairness by disable the scheduler tickling, and sorting the vcpu by the value of the credit remained(Please see the relevant ppt for VEE08). I have evaluated the credit scheduler of xen-3.1.0, it does agree with the option in the paper. > > However, when I evaluating the credit scheduler of xen-3.2.0, the results show that the scheduler can ensure the i/o fairness very well. I diff the sched_credit.c of the two versions of xen, they are the same. It seems that the credit scheduler of xen-3.2.0 does not sort the vcpu by credit or disable the scheduler tickling. So I wonder that what improvements xen have done to solove this problem? > > The experiment I have done is as follows: > The system has an AMD Opteron 848 processor, 4G physical memory, and a Gigabit ethernet NIC. I run eight xen VMs, and run netperf benchmark separatly. Every VM has almost the same bandwidth of 120 Mb/s. > > Thank you very much for your help. > > Xiang ZHANG > National Research Center for Intelligent Computing Systems > Institute of Computing Technology, Chinese Academy of Sciences, Beijing, > P.R.China_______________________________________________ Xen-devel mailing list Xen-devel@lists.xensource.com http://lists.xensource.com/xen-devel
Atsushi SAKAI
2008-May-26 06:50 UTC
Re: Re: [Xen-devel] Why the credit scheduler of xen-3.2.0 can ensure theI/O fairness?
Hi, Xiang tickling itself exists in first version of credit scheduler. changeset: New weighted fair-share CPU scheduler w/ automatic SMP load balancing changeset 10168: e539abd27a0f http://xenbits.xensource.com/xen-unstable.hg?rev=10168 Please check Table.1 and Section4.2 of your citation. Thanks Atsushi SAKAI "zhangxiang" <zhangxiang@ncic.ac.cn> wrote:> Hello Atsushi: > > Thank you very much for your reply! > I have already read that patch. In the paper "Scheduling I/O in Virtual Machine Monitors", the author said that fixing event channel notification can only ensure the I/O fairness with the tickling scheduler disable, and the scheduler tickling is the culprit of the I/O unfairness even the event channel code is fixed (Please see the Section 6.1). Although Xen-3.2.0 have fixxed the event channel code, it does not disable the scheduler tickling. So I think there should be something else xen-3.2.0 has done. > Thank you again for your help. > > Xiang ZHANG > National Research Center for Intelligent Computing Systems > Institute of Computing Technology, Chinese Academy of Sciences, Beijing, > P.R.China > > > > > > キ「シネヒ」コ Atsushi SAKAI > キ「ヒヘハアシ茱コ 2008-05-26 09:56:35 > ハユシネヒ」コ zhangxiang > ウュヒヘ」コ xen-devel > ヨ皮懃瘤搶鼈阨趙黼肬跛阯蜴鞜莅竏瘤艱黼會閭纉纐緕竏瘤鈬踈鈿肅竅闔蜴鳫筬鳫砠闥粤鬯竏瘤艱黼害敢傑絮禧箘過霞蔗痕緕砠緕齒竇闕蜴賢隨蒟崇纐聽宴銀芦甓壽瘤謫藻齦鼈啻冒葹鈑瘤腆綬葹鈑瘤臉釿蜒祟竡阡綺斐跛鍖纐纈閼葹鱚竇銓踟鱚痲鞜鞳竏繖蜴逼蜴帷鶯浴竏蜴酪鉗鴦↑蜒黶蜆癆竰繖蜚黹蒹糒跂閹竅鈿緕齦鱚薇聲蜥鈬齠跂癈黼鍋苺鱚蜴鞜鞳鬧壽癜闥黶蜆癆竅痺蓍纐薇聲蜥鈬齠磔粡黶碎黹蒹糒跂站跚鈑瘤齒鶯蜴頤磔跿閹竰繖蜚鱚轣蜴繖跂癈黼鱚跂銓韶肬崚徹県葹纐瘡竰繖蜚黹蒹糒跂閹遶凱窺穐蜚粹纉瘍鱚關闔蜴鞜鞳鬯被鬪緕纐瘡鈑竰繖蜚黹蒹糒跂閹遶凱荻穐鱚齦踉鼈阯癆黹蒹糒跂竅緕齦鱚薇聲蜥鈬齠鴒跛粡聿黹蒹籖竰繖蜚閹鴦蜿銖閹遯纖癇黶辣部黼纃癆竰繖蜚黹蒹糒跂閹遶凱荻粹纉鈿齒鶯頤磔竰繖蜚闥粡黶碎黹蒹糒跂站跚鈑囂鈔纈癆癆蜊頏阮纃緕葹粹鈬齒跫蜩頏閧跂躾壽纔鞳鱸辣銓葹粹鈬蜩癈肬跛阯鷓壽齷齡纃葹瘤鼠鳫鹸頏閭纉齒鬪看韆蜒瘡辣迴鴒瘤秤艨砠續蒹鴈續良樽鴣繪艾嵳鵺瘤鴣鈬纈矼釿蓁癇黼鞜鱇当纈嵳葹瘡迴齡黶辣矚鈔糘閹渦翌壽瘤鴒逡竏肬蒹跟懃瘤敞僧令闔瘡吶黼癇竏着銓纈肬侮跛蜃緕衷逅蜴囮齡纃侮齡蜚閹衷逅蜴壹竏鈿跫苹宙蜴纉巣痲纃閹噬蜈釿纉代蜉蜴膃豫勳宙蜴烝烝烝烝烝烝烝烝烝烝烝烝烝烝烝烝烝烝烝烝烝烝烝懊遶粤轣蛹蜴跚齡懊遶粤貔跚齡鶤銖阨鱆絎竢蔗痕蜩緕齒竇闕緕纐繻