From: Ming Z. <mi...@el...> - 2006-10-31 02:22:59
|
On Mon, 2006-10-30 at 19:18 -0700, Liang Yang wrote: > Hi Ming, > > Running multiple managers (each manager has maximum 2 workers) is a > workaround to solve the serialized outstanding I/O issue for IOMeter using > O_DIRECT.However, this workaround does get cache involved for both > sequential and random operations. ic. > > So I still prefer removing O_DIRECT. currently, u can remove it via modifying the source code directly. i will add this as run time option later. > > The reason I post here is I still think IOMeter shoud broadcast unique > random number to different managers no matter they are running on the same > machine or different machines. agree, at least for disk target. ps, did u fine time to test the patch for that cpu #? > > Regards, > > Liang > > ----- Original Message ----- > From: "Ming Zhang" <mi...@el...> > To: "Liang Yang" <mul...@ho...> > Cc: <iom...@li...>; > <iom...@li...> > Sent: Monday, October 30, 2006 6:25 PM > Subject: Re: [Iometer-user] IOMeter Workers under different managers usethe > same location for Random read/write. > > > > On Mon, 2006-10-30 at 16:00 -0700, Liang Yang wrote: > >> Hi, > >> > >> IOMeter originally plan to use multiple managers for multiple machines > >> (each > >> machine get its own manager). However, I do find some problems when I > >> spawned multiple managers on the same machine (multiple instances of > >> dynamo > >> are spawned on the same Linux platform). > > > > why u want to run multiple dynamo on same box? > > > >> > >> It seems those workers from different managers all use the same position > >> for > >> random read/write, this bring up cache effect even for random read/write > >> after comparing the data between single manager and multiple managers > >> with > >> the same number of I/Os(cache effect for sequential read/write with > >> multiple > >> workers is obvious as they re-read/re-write the same address). > > > > yes, quiet possible. > > > >> > >> I think the same random number icould be broadcasted to all managers > >> instead > >> of broadcasting a unique random number to each manager. This may not be a > >> good design even though people typically think managers should be on > >> different machines. > > > > it could be changed to start from different seeds. > > > >> > >> Can anyone give me some thoughts here? > >> > >> Thanks, > >> > >> Liang > >> > >> > >> ------------------------------------------------------------------------- > >> Using Tomcat but need to do more? Need to support web services, security? > >> Get stuff done quickly with pre-integrated technology to make your job > >> easier > >> Download IBM WebSphere Application Server v.1.0.1 based on Apache > >> Geronimo > >> http://sel.as-us.falkag.net/sel?cmd=lnk&kid=120709&bid=263057&dat=121642 > >> _______________________________________________ > >> Iometer-user mailing list > >> Iom...@li... > >> https://lists.sourceforge.net/lists/listinfo/iometer-user > > > > > > > ------------------------------------------------------------------------- > Using Tomcat but need to do more? Need to support web services, security? > Get stuff done quickly with pre-integrated technology to make your job easier > Download IBM WebSphere Application Server v.1.0.1 based on Apache Geronimo > http://sel.as-us.falkag.net/sel?cmd=lnk&kid=120709&bid=263057&dat=121642 > _______________________________________________ Iometer-user mailing list Iom...@li... https://lists.sourceforge.net/lists/listinfo/iometer-user |