ssd time to make head positioning (latency): <0.1ms hd max time to make head positioning (latency): 10ms ssd rate of read: 270MB/s random/sequential read (excluding latency) check that ssd is BLOCK (4kb mostly) oriented hd rate of read? 130MB/s sequential read? check that hd is BIT oriented write rate? random/sequencial? with these answers we can make a simple 'time' model of read/write, per device (use of raid0 (/dev/md0) is a device!, raid1 too (/dev/md1), raid5 (/dev/md2) ,raid6 (/dev/md3)) any device have this variables... just make a model and use the model to optimize minimal time to execute write/read 2011/2/2 Roberto Spadim <roberto@xxxxxxxxxxxxx>: > pros against closest head: > since we can use raid1 with identical disks (buyed at same time, with > near serial numbers) we can have disks with same time to fail > using closest head, the more used disk, will fail first > failing first we have time to change it (while the second isn't as > used as first device) > > but, think about it... > it's like a write-mostly not? > > > 2011/2/2 Roberto Spadim <roberto@xxxxxxxxxxxxx>: >> check that, read balance is: >> time based >> closest head >> round robin >> algorithms >> >> plus.... >> failed device problem and write-mostly >> >> with time based we can drop write-mosty.... just make the time of that >> device very high >> >> >> 2011/2/2 Roberto Spadim <roberto@xxxxxxxxxxxxx>: >>> it's cpu/mem consuming if use a complex model, and less cpu/mem >>> consuming if use a single model >>> >>> another idea.... >>> many algorithm.... >>> >>> first execute time based >>> it selected a bug (failed) device >>> execute closest head >>> if selected a bug (failed) device >>> execute round robin >>> if selected a bug (failed) device >>> select first usable non write-mostly >>> if selected a bug (failed) device >>> select first usable write-mostly >>> if end of devices, stop md raid >>> >>> to make this, today... we need a read_algorithm at >>> /sys/block/md0/xxxxxx, to select what algorith to use, write algorithm >>> is based on raid being used.. raid0 make linear and stripe, raid1 make >>> mirror, there's no algorithm to use here... >>> we need some files at /sys/block/md0/xxx to manage 'devices' time >>> model (parameters) >>> we need a adaptive algorithm to update parameters and make it closest >>> possible to real model of 'devices' >>> a raid0 have global parameters, inside raid0 devices have per device parameters >>> a raid1 over raid0, should use raid0 parameters >>> raid0 over devices, should use devices parameters >>> >>> >>> 2011/2/2 Roberto Spadim <roberto@xxxxxxxxxxxxx>: >>>> time based: is the time to: >>>> HD:head positioning , SSD: time to send command to ROM chip >>>> HD:read/write time (disk speed - rpm), SSD: time to write/read (time >>>> to ssd rom chip receive bytes) >>>> that's time based >>>> >>>> what is fast por read? >>>> consider that time based must know that disk is doing a I/O and that >>>> you have a time to end, this time to end is another time in algorithm >>>> >>>> for example: >>>> NBD (network block device) >>>> time to send read message + time to send command to rom or head positioning >>>> read/write time: time to nbd server return the read/write bytes >>>> >>>> what algorithm should do? >>>> calculate all time or all mirrors, including time to end current >>>> request (if only one request could be processed, or if allow more than >>>> 1 request, the time spent to start our command) >>>> after all time calculated, select the minimal value/device >>>> >>>> that's time based >>>> it's not based on round robin >>>> it's not based on closest head >>>> it's based on device speed to: >>>> *(1)position head/send rom command >>>> *(2)read/write time (per total of bytes read/write) >>>> *(3)time to start out request command (if don't allow more than 1 >>>> request per time, don't have a device queue) >>>> >>>> the total time per device will tell us the best device to read >>>> if we mix, nbd + ssd + hdd (5000rpm) + hdd(7500rpm) + hdd(10000rpm) + >>>> hdd(15000rpm) >>>> we can get the best read time using this algorithm >>>> the problem? we must run a constante benchmark to get this values *(1) >>>> *(2) *(3) and calculate good values of time spent on each process >>>> >>>> resuming... whe need a model of each device (simple-constants or very >>>> complex-neural network?), and calculate time spent per device >>>> nice? >>>> >>>> >>>> 2011/2/2 Robin Hill <robin@xxxxxxxxxxxxxxx>: >>>>> On Tue Feb 01, 2011 at 09:12:11PM -0200, Roberto Spadim wrote: >>>>> >>>>>> but the best algorithm is time based (minimize time to access data) >>>>>> >>>>> And what do you think takes the time accessing the data? In a rotating >>>>> disk, it's moving the heads - that's why the current strategy is nearest >>>>> head. In an SSD there's no head movement, so access time should be the >>>>> same for accessing any data, making it pretty much irrelevant which >>>>> strategy is used. >>>>> >>>>> Cheers, >>>>> Robin >>>>> >>>> >>>> >>>> >>>> -- >>>> Roberto Spadim >>>> Spadim Technology / SPAEmpresarial >>>> >>> >>> >>> >>> -- >>> Roberto Spadim >>> Spadim Technology / SPAEmpresarial >>> >> >> >> >> -- >> Roberto Spadim >> Spadim Technology / SPAEmpresarial >> > > > > -- > Roberto Spadim > Spadim Technology / SPAEmpresarial > -- 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
