Multipath

This document teaches you how to set up multipathing services for data storage.

Introduction
Multipathing services, generally deployed in enterprise environments, provide a means for high performance, load-balanced, and fault-tolerant data storage either locally or via a storage area network (SAN). Multipathing facilitates a single storage device to be transparently accessed across one or more paths. For example, if there are two connections from a server Host Bus Adapter (HBA) to two Fibre Channel switches and then to a SAN, when the HBA module loads and scans the bus, it will read four paths to the SAN: the paths from the server HBA to and from each Fibre Channel switch and at the storage device. Taking advantage of this situation, Multipath allows you to make use of each path simultaneously or independently to ensure a constant and reliable connection to the data in storage. Multipath serves as a failover for all connections points in the event of losing one path making critical data always available due to redundancy in the design and implementation.

In the most basic sense, multipathing is made of two distinct parts:  and. Device Mapper is the first key element of this application. Administrators are probably familiar with Device Mapper from LVM, EVMS, dm-crypt, or in this case, Multipath. In short, working within the kernel space Device Mapper takes one block device such as (as all SAN based targets will be some type of SCSI device) and maps it to another device.

On a lower level, Device Mapper creates a virtual block device accepting all of the commands of a regular block device, but passes on the actual data to the real block device. As previously stated, the mapping process is all handled in the kernel space and not in user space.

Multipath Tools is a set of userspace tools that interacts with the Device Mapper tools and creates structures for device handling, implementing I/O multipathing at the OS level. In a typical SAN environment, you will have multiple paths to the same storage device: a fiber card (or two) on your server that connects to a switch which then connects to the actual storage itself (as in the scenario discussed above). So administrators could possibly see the same device one to four times in such a situation (each card will see the LUN twice, once for each path it has available to it). Thus, a single drive could be recognized as,  ,  , and. If you were to mount to , for instance, you would be going over the singular path from one fiber card to a switch and then to a port on the same storage device. If any of those points were to fail, you would lose your storage device suddenly and have to unmount and remount with another device.

Consequently, this scenario is not ideal as you are only using one out of the four possible paths. This is where the combination of Multipath tools and Device Mapper are beneficial. As already explained, Device Mapper creates virtual block devices and then passes information to the real block devices.

Installation and Tools
You need to emerge  and. On the disk, you want to find the. You can use  (provided by   ) to do this.

Where DEVICE is the sd device, the ID will come back with a. Replace  with  , and you will have the proper ID that you'll put into the multipath   in. More on this in the next chapter.

Configuring Gentoo for multipathing
To configure Gentoo for multipath, your kernel needs the following settings:

In the kernel menu config, make sure CONFIG_SCSI_MULTI_LUN=y is set to ensure the SCSI subsystem is able to probe all Logical Unit Numbers (LUNs) (This is recommended as you'll stop scanning after ID 0 if you have a device on an ID of  but not   and then on an ID of   . Simply, you'll get your device for ID   but not   .) or whichever device you need for SCSI, such as a QLogic 2400 card, which is in the SCSI low-level drivers area.

For a better understanding, consider the following scenarios:

There are three drives with IDs of 0,1,2. Without the "probe all LUNs" setting, you will see IDs 0,1,2 as sda,sdb,sdc - all devices are seen. If you delete the ID 1 drive. IDs 0,2 will still be seen. It might seem to make sense that you would see sda and sdb now (sdc would move to sdb as there is no device to fill it up). However, if you don't probe all LUNs, it will perform in the following manner:

Scenario 1: Without "probe all LUNs", the scan will start and ID 0 will be seen. ID 0 will be set to sda and then move to find ID 1. If ID 1 is not detected, scanning will stop and be considered complete having perceived to have scanned all devices even if there is a device on ID 2 or any other subsequent ID. Reboot for scenario two.

Scenario 2: If you have "probe all LUNs", the scan will start and detect ID 0. This ID will be assigned sda and will continue to detect the next device. If ID 1 is not detected, scanning will continue to find more devices. ID 2 will be located and assigned to be sdb. If no devices (IDs) are detected beyond that, scanning will be considered complete.

So, once you probe all LUNs, all devices will be recognized and assigned an ID in Multipath.

Architectural Overview
As part of Multipath Tools, there are priority groups filled with the devices mentioned earlier. After you have configured  and started it with  , you can list the groups via. The output will look like the following:

By default, it will pick the first priority group (the first top round-robin for the EVA_SAN2, for instance, being ). In this instance, due to round robin it will bounce back and forth. But if one path was to fail, it would push all information to the other path and continue. Only if all the devices in a path fail will it actually fail and go to the secondary priority group.

Typical Configuration
A typical Multipath configuration looks like the following:

A typical /etc/multipath.conf file

A typical multipath configuration utilizing an EVA_SAN where the device information is in the kernel information regarding SAN hardware detection would look like:

EVA_SAN configuration

Setting Up Your Own Configuration
The multipath configuration is fairly simple to accomplish because the only file that needs modification is.

To begin, set the polling interview to how often (in seconds) path checks will be performed to ensure that the path is alive and healthy.

selector will be set at.

prio_callout : This one can be quite important, and there are a number of different priorities for different devices, such as:


 * mpath_prio_alua
 * mpath_prio_emc
 * mpath_prio_hds_modular
 * mpath_prio_netapp
 * mpath_prio_tpc

path_grouping_policy has a few different options: failover, multibus, group_by_prio. will only have one disk per priority group. will put all devices into one priority group. is done by a "priority value." So routes that have the same priority value will be grouped together, the priority values being determined by the callout.

no_path_retry is set to as most people don't want data to fail to send at all. So, if all paths fail, for instance, the I/Os will queue up until the device returns and then sends everything again. Depending on your transfer, this can cause load issues.

rr_min_io are the number of I/Os to do per path before switching to the next I/Os in the same group. If and  were in the same group, rr_min_io would do 100 I/Os to  then do 100 to , bouncing back and forth. This is a setting to tweak for each instance to maximize performance because the data load and size of transfers/request vary by company. The default in the case is, but some may prefer a smaller number in order to switch ports more often, when possible.

user_friendly_names make it easier to see which device you are working with. For example, if you set user_friendly_names to, then you'll see WWID instead of EVA_SAN for your device.

Acknowledgements
We would like to thank the following authors and editors for their contributions to this guide:


 * tsunam
 * Matthew Summers
 * Richard Anderson
 * Steve Rucker
 * nightmorph