Hmm, Roberto, I think we are close to theoretical maximum with some of the raid1/raid10 stuff already. and my nose tells me that we can gain more by minimizing CPU usage. Or maybe using some threading for raid modules - they all run single-threaded. Best regards keld On Wed, Feb 02, 2011 at 06:28:27PM -0200, Roberto Spadim wrote: > before, this thread i put at this page: > https://bbs.archlinux.org/viewtopic.php?pid=887267 > to make this mail list with less emails > > 2011/2/2 Keld Jørn Simonsen <keld@xxxxxxxxxx>: > > Hmm, Roberto, where are the gains? > > it?s dificult to talk... NCQ and linux scheduler don?t help a mirror, > they help a single device > a new scheduler for mirrors can be done (round robin, closest head, others) > > > I think it is hard to make raid1 better than it is today. > i don?t think, since head, is just for hard disk (rotational) not for > solid state disks, let?s not talk about ssd, just hard disk? a raid > with 5000rpm and 10000rpm disk, we will have better i/o read with > 10000rpm ? we don?t know the model of i/o for that device, but > probally will be faster, but when it?s busy we could use 5000rpm... > that?s the point, just closest head don?t help, we need know what?s > the queue (list of i/o being processed) and the time to read the > current i/o > > > Normally the driver orders the reads to minimize head movement > > and loss with rotation latency. Where can we improve that? > > no way to improve it, it?s very good! but per hard disk, not per mirror > but since we know it?s busy we can use another mirror (another disk > with same information), that?s what i want > > > Also, what about conflicts with the elevator algorithm? > elevator are based on model of disk, think disk as: linux elevator + > NCQ + disks, the sum of three infomration give us time based > infomrations to select best device > maybe making complex code (per elevator) we could know the time spent > to execute it, but it?s a lot of work, > for the first model, lets think about parameters of our model (linux > elevator + ncq + disks) > a second version we could implement elevator algorithm time > calculation (network block device NBD, have a elevator? at server side > + tcp/ip stack at client and server side, right?) > > > There are several scheduling algorithms available, and each has > > its merits. Will your new scheme work against these? > > Or is your new scheme just another scheduling algorithm? > > it?s a scheduling for mirrors > round balance is a algorithm for mirror > closest head is a algorithm for mirror > my 'new' algorith will be for mirror (if anyone help me coding for > linux kernel hehehe, i didn?t coded for linux kernel yet, just for > user space) > > noop, deadline, cfq isn?t for mirror, these are for raid0 problem > (linear, stripe if you hard disk have more then one head on your hard > disk) > > > I think I learned that scheduling is per drive, not per file system. > yes, you learned right! =) > /dev/md0 (raid1) is a device with scheduling (closest head,round robin) > /dev/sda is a device with scheduling (noop, deadline, cfq, others) > /dev/sda1 is a device with scheduling (it send all i/o directly to /dev/sda) > > the new algorithm is just for mirrors (raid1), i dont remeber about > raid5,6 if they are mirror based too, if yes they could be optimized > with this algorithm too > > raid0 don?t have mirrors, but information is per device striped (not > for linear), that?s why it can be faster... can make parallel reads > > with closest head we can?t use best disk, we can use a single disk all > time if it?s head closer, maybe it?s not the fastest disk (that?s why > we implent the write-mostly, we don?t make they usable for read, just > for write or when mirror fail, but it?s not perfect for speed, a > better algorithm can be made, for identical disks, a round robin work > well, better than closest head if it?s a solid state disk) > ok on a high load, maybe closest mirror is better than this algorithm? > yes, if you just use hard disk, if you mix hard disk+solid > state+network block device +floppy disks+any other device, you don?t > have the best algorithm for i/o over mirrors > > > > and is it reading or writing or both? Normally we are dependant on the > > reading, as we cannot process data before we have read them. > > OTOH writing is less time critical, as nobody is waiting for it. > it must be implemented on write and read, write for just time > calculations, read for select the best mirror > for write we must write on all mirrors (sync write is better, async > isn?t power fail safe) > > > Or is it maximum thruput you want? > > Or a mix, given some restraints? > it?s the maximum performace = what?s the better strategy to spent less > time to execute current i/o, based on time to access disk, time to > read bytes, time to wait others i/o being executed > > that?s for mirror select, not for disks i/o > for disks we can use noop, deadline, cfq scheduller (for disks) > tcp/ip tweaks for network block device > > a model identification must execute to tell the mirror select > algorithm what?s the model of each device > model: time to read X bytes, time to move head, time to start a read, > time to write, time time time per byte per kb per units > calcule time and select the minimal value calculated as the device > (mirror) to execute our read > > > > > > best regards > > keld > > thanks keld > > sorry if i make email list very big > > > > -- > Roberto Spadim > Spadim Technology / SPAEmpresarial -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
