The Linux SCSI Target Wiki
Network RAID allows for two or more Target machines to become physically redundant in order to mask hardware or storage array failures.
Network RAID1 is fully supported by LIO, it is now fully integrated into the RTS Director.
A prototype of a Target/Initiator ("T/I") repeater node was built with DRBD volumes as described below. A T/I repeater node is a physical or virtual machine that is running both iSCSI Target and Initiator stacks. The T/I repeater node was implemented with Open-iSCSI and LIO Targets running in DomUs under Xen. The Xen DomUs VM for the Targets were used to ease development.
The setup can be ported into LIO-VM. For the Initiators, both Open-iSCSI and Core-iSCSI can be used. For a multi-OS T/I repeater node, local Host OS local iSCSI storage can be imported through a hypervisor into LIO-VM.
A Network RAID1 demo setup can be built with virtual machines. In an early example based on Xen, both Initiator and Target nodes were fully redundant. The early example contained four Xen paravirtualized machines (two Target VMs and two Initiator VMs with ext3/OCFS2) running across two physical dom0 machines with 2x socket 2x core x86_64 with 8 GB of memory. The two Network RAID1 client VMs had no local storage (other than a Xen block device for the root filesystem), and were accessing storage on LIO Targets through Open-iSCSI. On both of the Network RAID1 Target nodes, volumes were created on top of available SCSI block devices. On the primary Network RAID1 Target node, the RAID1 array was built with:
mdadm --create /dev/md0 --level=1 --raid-devices=2 -bitmap=internal /dev/LIO-NR1-Elements/NR1-Local-Element --write-mostly /dev/LIO-NR1-Elements/NR1-Remote-Element
In the example, the Network RAID1 volume on the Target is constructed with Linux MD RAID1, with an internal write intent bitmap and write mostly element flag. The use of an internal bitmap for tracking changed blocks allows failed Network RAID1 Primary and Secondary nodes to recover quickly in the face of node failure. The use of the write mostly element flag is used on the Primary's remote iSCSI volume which represents Secondary's local storage. This ensures that READ ops coming from frontend iSCSI initiators are issued to the Primary's local storage.
The resulting Network RAID1 array looks as follows:
[root@bbtest2 ~]# cat /proc/mdstat Personalities : [raid1] md0 : active raid1 dm-2 dm-3(W) 10477504 blocks [2/2] [UU] bitmap: 1/160 pages [4KB], 32KB chunk unused devices: <none>
From there, a new volume group (LIO-NR1-VOL) is created and a new volume (NR1-PRIMARY-VOL) on the LIO-NR1 array (/dev/md0).
[root@bbtest2 ~]# lvs -v Finding all logical volumes LV VG #Seg Attr LSize Maj Min KMaj KMin Origin Snap% Move Copy% Log LV UUID NR1-Local-Element LIO-NR1-Elements 1 -wimao 10.00G 253 2 253 2 Qu7YhW-vdWo-IZPd-yDxP-sEbm-xM8L-y96RPD NR1-Remote-Element LIO-NR1-Elements 1 -wimao 9.99G 253 3 253 3 EEQewk-dhCW-UoMY-LgIK-QV8C-5Zlx-0Hppxc NR1-PRIMARY-VOL LIO-NR1-VOL 1 -wimao 9.98G 253 4 253 4 JYElqI-kJOD-QwRo-A68B-s6X9-jw6g-Jfyy1p LogVol00 VolGroup00 1 -wi-ao 3.75G -1 -1 253 0 69PKY5-5nIM-7TZX-vjoh-4sRJ-pALn-QDzNTq LogVol01 VolGroup00 1 -wi-ao 1.00G -1 -1 253 1 UwLWHP-J3Iv-s03T-q1gk-nXbE-hvBd-If7hg0
These iSCSI volumes and LIO-NR1 volumes need to be accessable on boot by LIO-Primary, and from there, the LVM UUID is passed into a virtual iBlock (BIO Sync Ack) or FILEIO (buffered Ack) in the LIO Target storage engine.
[root@bbtest2 ~]# target-ctl listluninfo tpgt=1 -----------------------------[LUN Info for iSCSI TPG 1]----------------------------- Status: ACTIVATED Execute/Left/Max Queue Depth: 0/32/32 SectorSize: 512 MaxSectors: 128 iBlock device: dm-4 LVM UUID: JYElqI-kJOD-QwRo-A68B-s6X9-jw6g-Jfyy1p Major: 253 Minor: 4 CLAIMED: IBLOCK Type: Direct-Access ANSI SCSI revision: 02 Unit Serial: JYElqI-kJOD-QwRo-A68B-s6X9-jw6g-Jfyy1p DIRECT EXPORTED iSCSI Host ID: 0 iSCSI LUN: 0 Active Cmds: 0 Total Bytes: 10716446720 ACLed iSCSI Initiator Node(s): iqn.1994-05.com.redhat:6e211fb3697 0 -> 0 iqn.1994-05.com.redhat:63a52b449156 0 -> 0
For testing purposes, all four VM disk images are located on iSCSI storage on their respective host virtualization machines. This storage is coming from one of the LIO Target nodes, and is MD RAID6 SATA with LVM2 on top of the array.
The prototype so far has proved very stable testing possible failure scenarios.
For production systems, we'd typically expect people to be using software or hardware RAID arrays, or Linux v2.6 lvm2 block devices.
What about growing the amount of LIO-NR1 storage available for frontend iSCSI initiators..?
There are at least two ways of doing this:
- ) Growing an existing LIO-NR1 volume (NR1-PRIMARY-VOL in the prototype) by building a new LIO-NR1 of local/remote storage objects. The frontend iSCSI initiators will have to rescan the logical unit for capacity, and then expand partition->filesystem.
- ) Creating a new LIO-NR1 array and volume and making a new iSCSI LUN available to frontend iSCSI initiators. These initiators can then create new filesystems or extend existing logical volumes.
Running Network RAID1 on Dom0 increases performance.
Using LVM volume block devices on the DomU Primary and Secondary T/I and VMs as elements of /dev/md0 on the LIO-NR1 machines seems to be a bit slower than raw SCSI block devices. We then create a LVM volume (
NR1-PRIMARY-VOL in the prototype) on top of /dev/md0 and this is the storage object that is exported to frontside iSCSI Initiators.
There is also a concern that using an internal write intent bitmap (which is pretty much a requirement for production) with MD has performance implications.
Having dedicated 1 Gb/sec or 10 Gb/sec ports between Network RAID1 nodes running jumbo frames for dedicated traffic on Dom0 should help improve latency and performance by reducing the number of interrupts produced by networking hardware.
Also, using dedicated CPU affinity for LIO Target threads on Dom0 is something that should be considered for production