Xen project Mailing List

Ian Campbell writes ("Re: [PATCH v6 00/21] libxl: domain save/restore: run in a separate process"):

> Does this mean this series is now ready to go in?

I think so. I'm just giving Shriram a chance to object.

I have no objections. I just finished testing the series with

xm (to ensure xend/remus was not broken). xl remus also fails over properly.

Things are good on that front.

But for the test case where I kill the backup (even with remote host replication), xl still crashes

the primary. [xend works properly in this case]. xl error output is at the end of the mail.

Either way, I have no objections to this series.

Tested-by: Shriram Rajagopalan <rshriram@xxxxxxxxx>

There is a more pressing matter that I just noticed.

The performance is extremely abysmal! Especially with xl.

Here is a comparative analysis between xend/remus and xl/remus

for a PV domU w/ and w/o suspend-event channel.

What is being measured ? time to suspend + time to resume

I am primarily concerned with the time to suspend and time to resume.

With event channel, this should be on the order of 1ms or so. With xenstore,

I would expect this to be max 5-7ms.

NB: This does not include the memcpy phase in xc_domain_save

Results: 32-bit PV domU w/ suspend event channel (2.6.32.2 xenolinux kernel)

xl-remus: ~1ms

xend-remus: ~1ms

So, for guests with suspend event channel support,

remus with xl/xend has the same suspend/resume overhead.

64-bit PV domU w/o suspend event channel (3.3.0 upstream kernel)

xl-remus: ~202ms !!!

xend-remus: ~2.2ms

This 202ms figure is same in both IanJ's tree and the baseline xen-unstable.

Looking back at the logs, this has been the same since January.

Is there some fixed timeout lurking in the code somewhere ?

====

xl error output, when killing backup VM (it crashes primary VM instead of resuming

it properly)

libxl: error: libxl_create.c:760:libxl__xc_domain_restore_done: restoring domain: Resource temporarily unavailable

libxl: error: libxl_create.c:844:domcreate_rebuild_done: cannot (re-)build domain: -3

libxl: error: libxl.c:1220:libxl_domain_destroy: non-existant domain 24

libxl: error: libxl_create.c:995:domcreate_complete: unable to destroy domain 24 following failed creation

migration target: Domain creation failed (code -3).

pagetables=2,cache_misses=0,emptypages=41

libxl: error: libxl_utils.c:363:libxl_read_exactly: file/stream truncated reading ipc msg header from domain 3 save/restore helper stdout pipe

libxl: error: libxl_exec.c:129:libxl_report_child_exitstatus: domain 3 save/restore helper [5620] died due to fatal signal Broken pipe

remus sender: libxl_domain_suspend failed (rc=-3)

Remus: Backup failed? resuming domain at primary.

thanks

shriram

Re: [Xen-devel] [PATCH v6 00/21] libxl: domain save/restore: run in a separate process