Re: [Qemu-devel] [PATCH v5 2/2] block/quorum: add simple read pattern support

[Liu Yuan]

On Mon, Aug 18, 2014 at 01:59:28PM +0800, Liu Yuan wrote:
> On Fri, Aug 15, 2014 at 03:59:04PM +0200, Benoît Canet wrote:
> > The Friday 15 Aug 2014 à 13:05:17 (+0800), Liu Yuan wrote :
> > > This patch adds single read pattern to quorum driver and quorum vote is 
> > > default
> > > pattern.
> > > 
> > > For now we do a quorum vote on all the reads, it is designed for 
> > > unreliable
> > > underlying storage such as non-redundant NFS to make sure data integrity 
> > > at the
> > > cost of the read performance.
> > > 
> > > For some use cases as following:
> > > 
> > >         VM
> > >   --------------
> > >   |            |
> > >   v            v
> > >   A            B
> > > 
> > > Both A and B has hardware raid storage to justify the data integrity on 
> > > its own.
> > > So it would help performance if we do a single read instead of on all the 
> > > nodes.
> > > Further, if we run VM on either of the storage node, we can make a local 
> > > read
> > > request for better performance.
> > > 
> > > This patch generalize the above 2 nodes case in the N nodes. That is,
> > > 
> > > vm -> write to all the N nodes, read just one of them. If single read 
> > > fails, we
> > > try to read next node in FIFO order specified by the startup command.
> > > 
> > > The 2 nodes case is very similar to DRBD[1] though lack of auto-sync
> > > functionality in the single device/node failure for now. But compared 
> > > with DRBD
> > > we still have some advantages over it:
> > > 
> > > - Suppose we have 20 VMs running on one(assume A) of 2 nodes' DRBD backed
> > > storage. And if A crashes, we need to restart all the VMs on node B. But 
> > > for
> > > practice case, we can't because B might not have enough resources to 
> > > setup 20 VMs
> > > at once. So if we run our 20 VMs with quorum driver, and scatter the 
> > > replicated
> > > images over the data center, we can very likely restart 20 VMs without any
> > > resource problem.
> > > 
> > > After all, I think we can build a more powerful replicated image 
> > > functionality
> > > on quorum and block jobs(block mirror) to meet various High Availibility 
> > > needs.
> > > 
> > > E.g, Enable single read pattern on 2 children,
> > > 
> > > -drive driver=quorum,children.0.file.filename=0.qcow2,\
> > > children.1.file.filename=1.qcow2,read-pattern=fifo,vote-threshold=1
> > > 
> > > [1] http://en.wikipedia.org/wiki/Distributed_Replicated_Block_Device
> > > 
> > > Cc: Benoit Canet <address@hidden>
> > > Cc: Eric Blake <address@hidden>
> > > Cc: Kevin Wolf <address@hidden>
> > > Cc: Stefan Hajnoczi <address@hidden>
> > > Signed-off-by: Liu Yuan <address@hidden>
> > > ---
> > >  block/quorum.c | 176 
> > > ++++++++++++++++++++++++++++++++++++++++++---------------
> > >  1 file changed, 129 insertions(+), 47 deletions(-)
> > > 
> > > diff --git a/block/quorum.c b/block/quorum.c
> > > index d5ee9c0..1235d7c 100644
> > > --- a/block/quorum.c
> > > +++ b/block/quorum.c
> > > @@ -24,6 +24,7 @@
> > >  #define QUORUM_OPT_VOTE_THRESHOLD "vote-threshold"
> > >  #define QUORUM_OPT_BLKVERIFY      "blkverify"
> > >  #define QUORUM_OPT_REWRITE        "rewrite-corrupted"
> > > +#define QUORUM_OPT_READ_PATTERN   "read-pattern"
> > >  
> > >  /* This union holds a vote hash value */
> > >  typedef union QuorumVoteValue {
> > > @@ -74,6 +75,8 @@ typedef struct BDRVQuorumState {
> > >      bool rewrite_corrupted;/* true if the driver must rewrite-on-read 
> > > corrupted
> > >                              * block if Quorum is reached.
> > >                              */
> > > +
> > > +    QuorumReadPattern read_pattern;
> > >  } BDRVQuorumState;
> > >  
> > >  typedef struct QuorumAIOCB QuorumAIOCB;
> > > @@ -117,6 +120,7 @@ struct QuorumAIOCB {
> > >  
> > >      bool is_read;
> > >      int vote_ret;
> > > +    int child_iter;             /* which child to read in fifo pattern */
> > 
> > I don't understand what "fifo pattern" could mean for a bunch of disk
> > as they are not forming a queue.
> 
> Naming isn't 100% accurate but as in Eric's comment (see below), both FIFO and
> Round-Robin can be used for two different patterns.
> 
> > Maybe round-robin is more suitable but your code does not implement
> > round-robin since it will alway start from the first disk.
> > 
> > Your code is scanning the disks set it's a scan pattern.
> > 
> > That said is it a problem that the first disk will be accessed more often 
> > than the other ?
> 
> As my commit log documented, the purpose of the read pattern I added is to
> speed up read against quorum original read pattern. And the use case is clear
> (I hope so) and you can take DRBD as a good example for why we need it. Of
> course we are far away from DRBD, which need a recovery logic after all kinds 
> of
> failures. My patch set can be taken as a prelimitary step to implement a DRBD
> like service driver.
> 
> Eric previously commented on two read patterns that might be useful:
> 
> "Should we offer multiple modes in addition to 'quorum'?  For example, I
> could see a difference between 'fifo' (favor read from the first quorum
> member always, unless it fails, good when the first member is local and
> other member is remote) and 'round-robin' (evenly distribute reads; each
> read goes to the next available quorum member, good when all members are
> equally distant)."
> 
> > You will have to care to insert disks in different order on each QEMU to 
> > spread the load.
> 
> This is another use case that my patch set didn't try to solve.
> 
> > Shouldn't the code try to spread the load by circling on the disk like a 
> > real round robin pattern ?
> 
> Probably not on my patch set, but we can add a yet another round robin pattern
> if anyone is intrested.

Benoit, ping...