Q re sync_completed

Q re sync_completed

[Eyal Lebedinsky]

I have scripts that do a raid check, then proceed to identify any files
affected. I then manually deal with these.

I have a few issues with this RADI6 setup, here is one.

I am setting sync_min and sync_max, start a check and wait for sync_completed
to equal sync_max.

I assumed that when equal it means that this address was "completed". After
doing this for a while I observed that this is probably not the case.

My expectation is that sync_completed has 'none' until it finished a chunk.
It then updates it with later completed ones. When it reaches sync_max
it pauses, and I then raise sync_max for the next area. This way I can
tell where a mismatch occurs. If sync_completed is set before a chunk
is completed then I may fetch mismatch_cnt too early (while the last
chunk is still being checked). This seems to be the case.

Q: Is this the case?

Setting ranges that are too small (minimum is 1024) makes the check
*very* slow. I notice that ranges of 1m or even 4m are required to
get the check to move along close to the maximum speed.

Q: Does the check take time to speed up rather than immediately go at
the nominated sync_speed_max rate?

TIA

-- 
Eyal Lebedinsky	(eyal@eyal.emu.id.au)

Re: Q re sync_completed

[NeilBrown]

On Sun, 13 Feb 2011 11:01:43 +1100 Eyal Lebedinsky <eyal@eyal.emu.id.au>
wrote:

> I have scripts that do a raid check, then proceed to identify any files
> affected. I then manually deal with these.
> 
> I have a few issues with this RADI6 setup, here is one.
> 
> I am setting sync_min and sync_max, start a check and wait for sync_completed
> to equal sync_max.
> 
> I assumed that when equal it means that this address was "completed". After
> doing this for a while I observed that this is probably not the case.
> 
> My expectation is that sync_completed has 'none' until it finished a chunk.
> It then updates it with later completed ones. When it reaches sync_max
> it pauses, and I then raise sync_max for the next area. This way I can
> tell where a mismatch occurs. If sync_completed is set before a chunk
> is completed then I may fetch mismatch_cnt too early (while the last
> chunk is still being checked). This seems to be the case.
> 
> Q: Is this the case?

The intention of sync_completed is that it is only updated after
sync/check/repair/recovery has actually completed to that point.  It may be
updated well *after* the sync has happened, but should never be updated
*before*.

However it is entirely possible that the code is not 100% correct.
If you give me details of what you are seeing together with precise kernel
version number I can try to explain them for you.


> 
> Setting ranges that are too small (minimum is 1024) makes the check
> *very* slow. I notice that ranges of 1m or even 4m are required to
> get the check to move along close to the maximum speed.
> 
> Q: Does the check take time to speed up rather than immediately go at
> the nominated sync_speed_max rate?

This is almost certainly an artifact of the way disk drives work.

To get streaming reads from a disk drive you need to request at least a whole
cylinder at a time.  As cylinders differ in size, it really only works if you
request multiple cylinders at a time.

I don't know how big cylinders are these days but I suspect they are a few
hundred K to a Meg.  So needing 4M at a time to get streaming happening
doesn't surprise me at all.

NeilBrown

Re: Q re sync_completed

[Eyal Lebedinsky]

My bad, my script has the wrong access order
	set `cat $sys/sync_action $sys/mismatch_cnt $sys/sync_completed`
mismatch_cnt should be fetched last.

cheers
	Eyal

On 02/13/11 15:32, NeilBrown wrote:
> On Sun, 13 Feb 2011 11:01:43 +1100 Eyal Lebedinsky<eyal@eyal.emu.id.au>
> wrote:
>
>> I have scripts that do a raid check, then proceed to identify any files
>> affected. I then manually deal with these.
>>
>> I have a few issues with this RADI6 setup, here is one.
>>
>> I am setting sync_min and sync_max, start a check and wait for sync_completed
>> to equal sync_max.
>>
>> I assumed that when equal it means that this address was "completed". After
>> doing this for a while I observed that this is probably not the case.
>>
>> My expectation is that sync_completed has 'none' until it finished a chunk.
>> It then updates it with later completed ones. When it reaches sync_max
>> it pauses, and I then raise sync_max for the next area. This way I can
>> tell where a mismatch occurs. If sync_completed is set before a chunk
>> is completed then I may fetch mismatch_cnt too early (while the last
>> chunk is still being checked). This seems to be the case.
>>
>> Q: Is this the case?
>
> The intention of sync_completed is that it is only updated after
> sync/check/repair/recovery has actually completed to that point.  It may be
> updated well *after* the sync has happened, but should never be updated
> *before*.
>
> However it is entirely possible that the code is not 100% correct.
> If you give me details of what you are seeing together with precise kernel
> version number I can try to explain them for you.
>
>> Setting ranges that are too small (minimum is 1024) makes the check
>> *very* slow. I notice that ranges of 1m or even 4m are required to
>> get the check to move along close to the maximum speed.
>>
>> Q: Does the check take time to speed up rather than immediately go at
>> the nominated sync_speed_max rate?
>
> This is almost certainly an artifact of the way disk drives work.
>
> To get streaming reads from a disk drive you need to request at least a whole
> cylinder at a time.  As cylinders differ in size, it really only works if you
> request multiple cylinders at a time.
>
> I don't know how big cylinders are these days but I suspect they are a few
> hundred K to a Meg.  So needing 4M at a time to get streaming happening
> doesn't surprise me at all.
>
> NeilBrown

-- 
Eyal Lebedinsky	(eyal@eyal.emu.id.au)