Re: [Qemu-devel] [PATCH V5 0/9] nvdimm: guarantee persistence of QEMU writes to persistent memory

[Dr. David Alan Gilbert]

* Junyan He (address@hidden) wrote:
> > Also, there was a discussion about leaving the code unchanged but adding
> > an nvdimm_flush() call at the very end of migration.  I think someone
> > benchmarked it but can't find the email.  Please post a link or
> > summarize the results, because that approach would be much less
> > invasive.  Thanks!
> 
> 
> And previous comments:
> 
> 
> > > > 2. The migration/ram code is invasive.  Is it really necessary to
> > > >    persist data each time pages are loaded from a migration stream?  It
> > > >    seems simpler to migrate as normal and call pmem_persist() just once
> > > >    after RAM has been migrated but before the migration completes.
> > >
> > > The concern is about the overhead of cache flush.
> > >
> > > In this patch series, if possible, QEMU will use pmem_mem{set,cpy}_nodrain
> > > APIs to copy NVDIMM blocks. Those APIs use movnt (if it's available) and
> > > can avoid the subsequent cache flush.
> > >
> > > Anyway, I'll make some microbenchmark to check which one will be better.
> 
> > The problem is not just the overhead; the problem is the code
> > complexity; this series makes all the paths through the migration code
> > more complex in places we wouldn't expect to change.
> 
> I already use the migration info tool and list the result in the Mail just 
> after this patch set sent:
> 
> Disable all haozhong's pmem_drain and  pmem_memset_nodrain  kind function call
> and make the cleanup function do the flush job like this:
> 
> static int ram_load_cleanup(void *opaque)
> {
>     RAMBlock *rb;
>     RAMBLOCK_FOREACH(rb) {
>         if (ramblock_is_pmem(rb)) {
>             pmem_persist(rb->host, rb->used_length);
>          }
>     }
> 
>     xbzrle_load_cleanup();
>     compress_threads_load_cleanup();
> 
>     RAMBLOCK_FOREACH(rb) {
>         g_free(rb->receivedmap);
>         rb->receivedmap = NULL;
>     }
>     return 0;
> }
> 
> 
> The migrate info result is:
> 
> Haozhong's Manner
> 
> (qemu) migrate -d tcp:localhost:4444
> (qemu) info migrate
> globals:
> store-global-state: on
> only-migratable: off
> send-configuration: on
> send-section-footer: on
> capabilities: xbzrle: off rdma-pin-all: off auto-converge: off zero-blocks: 
> off compress: off events: off postcopy-ram: off x-colo: off release-ram: off 
> block: off return-path: off pause-before-switchover: off x-multifd: off 
> dirty-bitmaps: off postcopy-blocktime: off
> Migration status: completed
> total time: 333668 milliseconds
> downtime: 17 milliseconds
> setup: 50 milliseconds
> transferred ram: 10938039 kbytes
> throughput: 268.55 mbps
> remaining ram: 0 kbytes
> total ram: 11027272 kbytes
> duplicate: 35533 pages
> skipped: 0 pages
> normal: 2729095 pages
> normal bytes: 10916380 kbytes
> dirty sync count: 4
> page size: 4 kbytes
> (qemu)
> 
> 
> flush before complete
> 
> QEMU 2.12.50 monitor - type 'help' for more information
> (qemu) migrate -d tcp:localhost:4444
> (qemu) info migrate
> globals:
> store-global-state: on
> only-migratable: off
> send-configuration: on
> send-section-footer: on
> capabilities: xbzrle: off rdma-pin-all: off auto-converge: off zero-blocks: 
> off compress: off events: off postcopy-ram: off x-colo: off release-ram: off 
> block: off return-path: off pause-before-switchover: off x-multifd: off 
> dirty-bitmaps: off postcopy-blocktime: off
> Migration status: completed
> total time: 334836 milliseconds
> downtime: 17 milliseconds
> setup: 49 milliseconds
> transferred ram: 10978886 kbytes
> throughput: 268.62 mbps
> remaining ram: 0 kbytes
> total ram: 11027272 kbytes
> duplicate: 23149 pages
> skipped: 0 pages
> normal: 2739314 pages
> normal bytes: 10957256 kbytes
> dirty sync count: 4
> page size: 4 kbytes
> (qemu)
> 
> 
> So Haozhong's manner seems to be a little faster and I choose to keep that.
> 
> If you want to choose this manner, the code will be clean and no need for
> 
> >   typedef struct {
> >       void (*memset)(void *s, int c, size_t n);
> >       void (*memcpy)(void *dest, const void *src, size_t n);
> >   } MemoryOperations;
> 
> 
> performance is close, and I am a little new in Qemu:), so both options are OK 
> for me,
> 
> Which one do you prefer to?

The one with the least impact;  the migration code is getting more and
more complex, so having to do the 'if (is_pmem)' check everywhere isn't
nice,  passing an 'ops' pointer in is better.  However if you can do the
'flush before complete' instead then the amount of code change is a LOT
smaller.
The only other question is whether from your pmem view, the
flush-before-complete causes any problems;  in the worst case, how long
could the flush take?

Dave

> ________________________________
> From: Stefan Hajnoczi <address@hidden>
> Sent: Thursday, May 31, 2018 1:18:58 PM
> To: address@hidden
> Cc: address@hidden; Haozhong Zhang; address@hidden; address@hidden; 
> address@hidden; address@hidden; address@hidden; address@hidden; Junyan He; 
> address@hidden; address@hidden; address@hidden; address@hidden
> Subject: Re: [Qemu-devel] [PATCH V5 0/9] nvdimm: guarantee persistence of 
> QEMU writes to persistent memory
> 
> David Gilbert previously suggested a memory access interface.  I guess
> it would look something like this:
> 
>   typedef struct {
>       void (*memset)(void *s, int c, size_t n);
>       void (*memcpy)(void *dest, const void *src, size_t n);
>   } MemoryOperations;
> 
> That way code doesn't need if (pmem) A else B.  It can just do
> mem_ops->foo().  Have you looked into this idea?
> 
> Also, there was a discussion about leaving the code unchanged but adding
> an nvdimm_flush() call at the very end of migration.  I think someone
> benchmarked it but can't find the email.  Please post a link or
> summarize the results, because that approach would be much less
> invasive.  Thanks!
> 
> Stefan
--
Dr. David Alan Gilbert / address@hidden / Manchester, UK