Currently, the farm driver has a fatal problem in object recovery; it
blocks I/O requests long time while moving stale objects to the
backend store. Here is a test script to show that.
==
sheep -d /store/0 -z 0 -p 7000
sleep 2
collie cluster format -c 1 -b $1
collie vdi create test 100G -P
sheep
On 08/23/2012 10:58 PM, MORITA Kazutaka wrote:
> To fix this problem, we need to move stale objects in the worker
> thread. I've spent several days trying to fix it, but it seems to
> need a lot of work, and it looks difficult to stable the change in a
> short time.
>
Last time your patches to f
At Fri, 24 Aug 2012 00:03:09 +0800,
Liu Yuan wrote:
>
> On 08/23/2012 10:58 PM, MORITA Kazutaka wrote:
> > To fix this problem, we need to move stale objects in the worker
> > thread. I've spent several days trying to fix it, but it seems to
> > need a lot of work, and it looks difficult to stabl
At Fri, 24 Aug 2012 00:03:09 +0800,
Liu Yuan wrote:
> > $ ./farm.sh simple
> > using backend simple store
> > NameIdSizeUsed SharedCreation time VDI id Tag
> > test 1 100 GB 100 GB 0.0 MB 2012-08-23 11:29 7c2b25
> >
> > real0m1.016s
> > user0m0.092s
On 08/24/2012 01:06 AM, MORITA Kazutaka wrote:
> The result of simple store is not 1 minute but 1 second.
Oops, I indeed mistook it.
Thanks,
Yuan
--
sheepdog mailing list
sheepdog@lists.wpkg.org
http://lists.wpkg.org/mailman/listinfo/sheepdog
On 08/23/2012 10:58 PM, MORITA Kazutaka wrote:
> This introduces a storage driver 'simple_store' based on the current
> storage interface. The design of the new simple_store is similar to
> one of the farm driver. The main difference is that farm uses the
> sha1 based backend store for stale obje
At Fri, 24 Aug 2012 12:44:05 +0800,
Liu Yuan wrote:
>
> I am also in favor of simple implementation. I think a dedicated .stale
> directory idea could also apply to Farm. So How about following approach:
>
> merge simple store into Farm, then it actually run the code as
> current simple s
On 08/24/2012 01:37 PM, MORITA Kazutaka wrote:
> It is a good idea to refine the farm backend store too, but simple
> store should be added as another driver, I think. The simple driver
> would become a good tutorial to introduce new drivers.
Okay, I'd suggest use default_xxx as prefix instead of
At Fri, 24 Aug 2012 15:04:56 +0800,
Liu Yuan wrote:
>
> On 08/24/2012 01:37 PM, MORITA Kazutaka wrote:
> > It is a good idea to refine the farm backend store too, but simple
> > store should be added as another driver, I think. The simple driver
> > would become a good tutorial to introduce new d
On 08/24/2012 03:15 PM, MORITA Kazutaka wrote:
> At Fri, 24 Aug 2012 15:04:56 +0800,
> Liu Yuan wrote:
>>
>> On 08/24/2012 01:37 PM, MORITA Kazutaka wrote:
>>> It is a good idea to refine the farm backend store too, but simple
>>> store should be added as another driver, I think. The simple driver
On 08/24/2012 03:28 PM, Liu Yuan wrote:
> which is also relevant to
which is also not relevant to
--
sheepdog mailing list
sheepdog@lists.wpkg.org
http://lists.wpkg.org/mailman/listinfo/sheepdog
On 08/24/2012 03:28 PM, Liu Yuan wrote:
>>> >>
>>> >> I'll go to update the farm's io path to use defaul_xxx functions, I
>>> >> think this simple store implementation can be nicely applied to farm.
>> >
>> > The update will change the data layout of farm. How about introducing
>> > 'farm2'? I g
At Fri, 24 Aug 2012 15:28:07 +0800,
Liu Yuan wrote:
>
> I think simple_store.c itself won't cause confusion. 'default store'
> will instead cause greater confusion.
The reason I think 'simple' is not good is that we had a 'simple'
driver before which had a completely different data layout. There
On 08/24/2012 03:37 PM, MORITA Kazutaka wrote:
> The reason I think 'simple' is not good is that we had a 'simple'
> driver before which had a completely different data layout. There may
> be some users who still use a simple driver with old sheepdog version.
> If they upgrade to the latest versio
At Fri, 24 Aug 2012 15:40:14 +0800,
Liu Yuan wrote:
>
> On 08/24/2012 03:37 PM, MORITA Kazutaka wrote:
> > The reason I think 'simple' is not good is that we had a 'simple'
> > driver before which had a completely different data layout. There may
> > be some users who still use a simple driver wi
On 08/23/2012 10:58 PM, MORITA Kazutaka wrote:
> +static int simple_store_read_from_path(uint64_t oid, char *path,
> +struct siocb *iocb)
> +{
> + int flags = def_open_flags, fd, ret = SD_RES_SUCCESS;
> + ssize_t size;
> +
> + if (!is_data_obj(oid))
>
At Fri, 24 Aug 2012 15:57:37 +0800,
Liu Yuan wrote:
>
> On 08/23/2012 10:58 PM, MORITA Kazutaka wrote:
> > +static int simple_store_read_from_path(uint64_t oid, char *path,
> > + struct siocb *iocb)
> > +{
> > + int flags = def_open_flags, fd, ret = SD_RES_SUCCES
On 08/24/2012 04:08 PM, MORITA Kazutaka wrote:
> I think both farm and plain don't need it. Can you explain a scenario
> where it can cause a problem?
Qemu can issue multiple requests on the same oid, doesn't it? If there
is some kind of write barrier to prevent reading stale oid or actually
no p
At Fri, 24 Aug 2012 16:13:51 +0800,
Liu Yuan wrote:
>
> On 08/24/2012 04:08 PM, MORITA Kazutaka wrote:
> > I think both farm and plain don't need it. Can you explain a scenario
> > where it can cause a problem?
>
> Qemu can issue multiple requests on the same oid, doesn't it? If there
Yes, but
19 matches
Mail list logo