[ofa-general] SRP target/LVM HA configuration

[Vu Pham]

I would second to Stanley - client hosts 
fail-over/load-balancing would be straightforward

For linux host you have several tools: sw raid, lvm or 
dm-multipath

   -vu

> I had looked at this configuration as well and decided to use the volume
> management at the clients to mirror the data. Windows LDM mirrored
> across 2 SRPT servers and Linux md RAID 1 mirrored.
> 
> This provides transparent failover and the SRP client/host will rebuild
> the slices that went offline.
> 
>> -----Original Message-----
>> From: general-bounces at lists.openfabrics.org 
>> [mailto:general-bounces at lists.openfabrics.org] On Behalf Of 
>> Daniel Pocock
>> Sent: Tuesday, March 11, 2008 4:26 PM
>> To: general at lists.openfabrics.org
>> Subject: [ofa-general] SRP target/LVM HA configuration
>>
>>
>>
>>
>>
>>
>> I'm contemplating a HA configuration based on SRP and LVM (or 
>> maybe EVMS).
>>
>> There are many good resources based on NFS and drbd (see 
>> http://www.linux-ha.org/HaNFS) but it would be more flexible to work 
>> with block level (e.g SRP) rather than file level (NFS).  Obviously, 
>> SRP/RDMA offers a major performance benefit compared with drbd (which 
>> uses IP).
>>
>> Basically, I envisage the primary server having access to the 
>> secondary 
>> (passive) server's disk using SRP, and putting both the local 
>> (primary) 
>> disk and SRP (secondary) disk into RAID1.  The RAID1 set 
>> would contain a 
>> volume group and multiple volumes - which would, in turn, be 
>> SRP targets 
>> (for VMware to use) or possibly NFS shares.
>>
>> This leads me to a few issues:
>>
>> - Read operations - would it be better for the primary to 
>> read from both 
>> disks, or just it's own disk?  Using drbd, the secondary disk is not 
>> read unless the primary is down.  However, given the 
>> performance of SRP, 
>> I suspect that reading from both the local and SRP disk would give a 
>> boost to performance.
>>
>> - Does it make sense to use md or LVM to combine a local disk 
>> and an SRP 
>> disk into RAID1 (or potentially RAID5)?  Are there technical 
>> challenges 
>> there, given that one target is slightly faster than the other?
>>
>> - Fail-over - when the secondary detects that the primary is 
>> down, can 
>> it dynamically take the place of the failed SRP target?  Will the 
>> end-user initiators (e.g. VMWare, see diagram below) be confused when 
>> the changeover occurs?  Is there the possibility of data 
>> inconsistency 
>> if some write operations had been acknowledged by the primary but not 
>> propagated to the secondary's disk at the moment when the 
>> failure occurred?
>>
>> - Recovery - when the old primary comes back online as a 
>> secondary, it 
>> will need to resync it's disk - is a partial resync possible, 
>> or is full 
>> rebuild mandatory?
>>
>>
>> Diagram:
>>
>>
>> Disk--Primary Server-------------------SRP Initiator (e.g. VMware ESX)
>>         |                       +------NFS client     
>>         |                       .
>>        SRP                      .
>>    (RAID1 of primary's          .
>>    disk and secondary's         .
>>       disk)                     . (fail-over path to storage
>>         |                       .  when primary is down)
>> Disk--Secondary Server. . . . . .
>>
>>
>>
>> _______________________________________________
>> general mailing list
>> general at lists.openfabrics.org 
>> http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general
>>
>> To unsubscribe, please visit 
>> http://openib.org/mailman/listinfo/openib-general
>>
> _______________________________________________
> general mailing list
> general at lists.openfabrics.org
> http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general
> 
> To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general