Difference between revisions of "Mellanox-Cinder"
m |
m (→Overview) |
||
| Line 1: | Line 1: | ||
= Overview = | = Overview = | ||
| − | Mellanox added iSCSI Extensions for RDMA (iSER) that does not need the TCP layer and therefore provides lower latency.<br /> | + | Mellanox added iSCSI Extensions for RDMA (iSER) that does not need the TCP layer and therefore provides lower latency and higher performance.<br /> |
It permits data to be transferred directly into and out of SCSI computer memory buffers (which connects computers to storage devices) without intermediate data copies.<br /> | It permits data to be transferred directly into and out of SCSI computer memory buffers (which connects computers to storage devices) without intermediate data copies.<br /> | ||
iSER is a server network protocol that extends the Internet Small Computer System Interface (iSCSI) protocol to use Remote Direct Memory Access (RDMA).<br /><br /> | iSER is a server network protocol that extends the Internet Small Computer System Interface (iSCSI) protocol to use Remote Direct Memory Access (RDMA).<br /><br /> | ||
Revision as of 13:36, 31 December 2014
Overview
Mellanox added iSCSI Extensions for RDMA (iSER) that does not need the TCP layer and therefore provides lower latency and higher performance.
It permits data to be transferred directly into and out of SCSI computer memory buffers (which connects computers to storage devices) without intermediate data copies.
iSER is a server network protocol that extends the Internet Small Computer System Interface (iSCSI) protocol to use Remote Direct Memory Access (RDMA).
Please contact Mellanox for further information and benchmark results.
Installation
Provisioned through standard Cinder installation.
Configuration
In order to enable iSER, need to adjust these values at /etc/cinder/cinder.conf:
iser_ip_address = <ipoib/roce_address> volume_driver = cinder.volume.drivers.lvm.LVMISERDriver
iser_ip_address is required to do a "discovery" over the IB/RoCE interface from the initiator side.
volume_driver points cinder to use the ISERDriver, instead the LVMISCSIDriver.
On nova-compute side change the following in /etc/nova/nova.conf:
libvirt_volume_drivers = iser=nova.virt.libvirt.volume.LibvirtISERVolumeDriver
Restart Cinder and Nova compute services
#/etc/init.d/openstack-cinder-volume restart #/etc/init.d/openstack-nova-compute restart
For troubleshooting issues refer to section Known Issues
Known Issues
(1) “scsi-target-utils” package
Note: for RH6.4 or below: “scsi-target-utils” package needs to be 1.0.38 and up, if you use RDO installation you should check the version. The RPM can be found in http://www.mellanox.com/downloads/solutions/rpms/
#wget http://www.mellanox.com/downloads/solutions/rpms/scsi-target-utils-1.0.39-v1.0.39.c1135a.x86_64.rpm #rpm -Uvh scsi-target-utils-1.0.39-v1.0.39.c1135a.x86_64.rpm
In addition, RH6.5 has “scsi-target-utils” package with version 1.0.24-10, which is good enough, no further action is needed for RHEL6.5
(2) Boot Instance from volume fails when iSER is enabled
Steps to reproduce the bug:
- Create new volume with image source.
- Launch a new instance boot from the above volume using iSER transport.
The bug is defined here: [1]
A patch can be found in here: [2]
To apply the patch run the following commands on all nova-compute nodes:
# wget http://www.mellanox.com/downloads/solutions/openstack/havana/patches/volume.patch # patch -p1 < volume.patch # service nova-compute restart
(3) Flow control
In case the network is Ethernet, Make sure that flow control is enabled on the switches on the relevant ports across the network.
(4) lio support limitation.
Open-source implementation of SCSI target in Openstack can be managed by either TGT or LIO.
Currently in Openstack, iSER is supported over TGT only, and iSER support over LIO is scheduled for Kilo version.
In case LIO is the default SCSI target, one must one install TGT and alter Openstack Cinder conf file to use TGT instead of lio.
References
1. OpenStack solution page at Mellanox site
5. Cinder-node
Return to Mellanox-OpenStack wiki page.
For more details, please refer any inquiries to openstack@mellanox.com.
