HP StorageWorks SAN Virtualization Services Platform Best Practices Guide (5697-0935, May 2011)

Table 2 (page 41) show the recommendation from HP for the sizing of SVSP configurations that
use synchronous mirrors heavily. The term safe means that the performance of two DPMs is examined
and cut in half to account for failure of one DPM.
CAUTION: Existing installations may have relied on performance numbers available for basic
virtual disks that were much more optimistic than these measured for synchronous mirroring.
Table 2 Safe synchronous mirroring performance sizing
4 DPM groups3 DPM groups2 DPM groups1 DPM groupWorkload
857,280642,960428,640214,320512 B 100% read (IO/s)
700,064525,048350,032175,0168 KB 100% read (IO/s)
99,68874,76649,84424,9228 KB 100% write (IO/s)
208,768156,576104,38452,1928 KB 60% read/40% write (IO/s)
10,3447,7585,1722,586128 KB read (MB/s)
1,232924616308128 KB write (MB/s)
Operating modes
Operating modes for synchronous mirror tasks are:
• Always Synchronized—If one copy fails, all copies in the task are suspended. When the DPM
running the synchronous mirror task identifies the first write error on one of the copies, all
subsequent I/Os are rejected. The other copy in the task is suspended immediately and writes
to the primary virtual disk will fail. After resolving the issue, run the Resume command to
reinstate the mirroring tasks.
• Continue-on-Fail—If a copy fails, the host that is using the synchronous mirror can remain
active on the remaining copy and the other copy in the group starts tracking which chunks
were updated in the reallocation table metadata if the dirty flag is set (see the HP StorageWorks
SAN Virtualization Service Platform Administrator Guide). After resolving the issue and
resynchronizing the task, the reallocation tables of the other copies are used to determine
which chunks need to be written to the task's virtual disk to synchronize it. This avoids copying
the entire primary virtual disk to the secondary virtual disk. Normal journaling is active only
when the task has failed copies. If the dirty flag is not set, no tracking is done, and a full copy
is required to resynchronize the legs.
Synchronous mirror best practice
When creating synchronous mirrors, it is desirable to have the source and destination on different
storage pools or arrays. Having both source and destination (or multiple synchronous legs) in the
same pool does work, but could lead to overloading of the back-end storage and this lacks value
from an availability standpoint. Distributing the mirrors is consistent with the typical customer use
case for mirroring.
With synchronous mirrors, only the I/O in flight is at risk when a site goes down. The RPO for
synchronous mirrors is zero and the requirement of not losing data is satisfied. Because every host
write I/O is only acknowledged to the host, when all write destinations (the primary and all
secondaries) have successfully written and acknowledged the write, the slowest member of the
mirror group defines the maximum performance of the whole group. Should any write destination
experience latencies above normal Fibre Channel latencies due to distances or network issues,
write I/O numbers for this synchronous mirror group will be decreased. For this reason, synchronous
mirrors are only suited for remote destinations over short to medium distances (typically Fibre
Channel limits) and low latency networks.
Overview 41