swift/doc/source/deployment_guide.rst

Deployment Guide

This document provides general guidance for deploying and configuring Swift. Detailed descriptions of configuration options can be found in the configuration documentation <config/index>.

Hardware Considerations

Swift is designed to run on commodity hardware. At Rackspace, our storage servers are currently running fairly generic 4U servers with 24 2T SATA drives and 8 cores of processing power. RAID on the storage drives is not required and not recommended. Swift's disk usage pattern is the worst case possible for RAID, and performance degrades very quickly using RAID 5 or 6.

Deployment Options

The Swift services run completely autonomously, which provides for a lot of flexibility when architecting the hardware deployment for Swift. The 4 main services are:

1. Proxy Services
2. Object Services
3. Container Services
4. Account Services

The Proxy Services are more CPU and network I/O intensive. If you are using 10g networking to the proxy, or are terminating SSL traffic at the proxy, greater CPU power will be required.

The Object, Container, and Account Services (Storage Services) are more disk and network I/O intensive.

The easiest deployment is to install all services on each server. There is nothing wrong with doing this, as it scales each service out horizontally.

At Rackspace, we put the Proxy Services on their own servers and all of the Storage Services on the same server. This allows us to send 10g networking to the proxy and 1g to the storage servers, and keep load balancing to the proxies more manageable. Storage Services scale out horizontally as storage servers are added, and we can scale overall API throughput by adding more Proxies.

If you need more throughput to either Account or Container Services, they may each be deployed to their own servers. For example you might use faster (but more expensive) SAS or even SSD drives to get faster disk I/O to the databases.

A high-availability (HA) deployment of Swift requires that multiple proxy servers are deployed and requests are load-balanced between them. Each proxy server instance is stateless and able to respond to requests for the entire cluster.

Load balancing and network design is left as an exercise to the reader, but this is a very important part of the cluster, so time should be spent designing the network for a Swift cluster.

Web Front End Options

Swift comes with an integral web front end. However, it can also be deployed as a request processor of an Apache2 using mod_wsgi as described in Apache Deployment Guide <apache_deployment_guide>.

Preparing the Ring

The first step is to determine the number of partitions that will be in the ring. We recommend that there be a minimum of 100 partitions per drive to insure even distribution across the drives. A good starting point might be to figure out the maximum number of drives the cluster will contain, and then multiply by 100, and then round up to the nearest power of two.

For example, imagine we are building a cluster that will have no more than 5,000 drives. That would mean that we would have a total number of 500,000 partitions, which is pretty close to 2^19, rounded up.

It is also a good idea to keep the number of partitions small (relatively). The more partitions there are, the more work that has to be done by the replicators and other backend jobs and the more memory the rings consume in process. The goal is to find a good balance between small rings and maximum cluster size.

The next step is to determine the number of replicas to store of the data. Currently it is recommended to use 3 (as this is the only value that has been tested). The higher the number, the more storage that is used but the less likely you are to lose data.

It is also important to determine how many zones the cluster should have. It is recommended to start with a minimum of 5 zones. You can start with fewer, but our testing has shown that having at least five zones is optimal when failures occur. We also recommend trying to configure the zones at as high a level as possible to create as much isolation as possible. Some example things to take into consideration can include physical location, power availability, and network connectivity. For example, in a small cluster you might decide to split the zones up by cabinet, with each cabinet having its own power and network connectivity. The zone concept is very abstract, so feel free to use it in whatever way best isolates your data from failure. Each zone exists in a region.

A region is also an abstract concept that may be used to distinguish between geographically separated areas as well as can be used within same datacenter. Regions and zones are referenced by a positive integer.

You can now start building the ring with:

swift-ring-builder <builder_file> create <part_power> <replicas> <min_part_hours>

This will start the ring build process creating the <builder_file> with 2^<part_power> partitions. <min_part_hours> is the time in hours before a specific partition can be moved in succession (24 is a good value for this).

Devices can be added to the ring with:

swift-ring-builder <builder_file> add r<region>z<zone>-<ip>:<port>/<device_name>_<meta> <weight>

This will add a device to the ring where <builder_file> is the name of the builder file that was created previously, <region> is the number of the region the zone is in, <zone> is the number of the zone this device is in, <ip> is the ip address of the server the device is in, <port> is the port number that the server is running on, <device_name> is the name of the device on the server (for example: sdb1), <meta> is a string of metadata for the device (optional), and <weight> is a float weight that determines how many partitions are put on the device relative to the rest of the devices in the cluster (a good starting point is 100.0 x TB on the drive).Add each device that will be initially in the cluster.

Once all of the devices are added to the ring, run:

swift-ring-builder <builder_file> rebalance

This will distribute the partitions across the drives in the ring. It is important whenever making changes to the ring to make all the changes required before running rebalance. This will ensure that the ring stays as balanced as possible, and as few partitions are moved as possible.

The above process should be done to make a ring for each storage service (Account, Container and Object). The builder files will be needed in future changes to the ring, so it is very important that these be kept and backed up. The resulting .tar.gz ring file should be pushed to all of the servers in the cluster. For more information about building rings, running swift-ring-builder with no options will display help text with available commands and options. More information on how the ring works internally can be found in the Ring Overview <overview_ring>.

Running object-servers Per Disk

The lack of true asynchronous file I/O on Linux leaves the object-server workers vulnerable to misbehaving disks. Because any object-server worker can service a request for any disk, and a slow I/O request blocks the eventlet hub, a single slow disk can impair an entire storage node. This also prevents object servers from fully utilizing all their disks during heavy load.

Another way to get full I/O isolation is to give each disk on a storage node a different port in the storage policy rings. Then set the servers_per_port <object-server-default-options> option in the object-server config. NOTE: while the purpose of this config setting is to run one or more object-server worker processes per disk, the implementation just runs object-servers per unique port of local devices in the rings. The deployer must combine this option with appropriately-configured rings to benefit from this feature.

Here's an example (abbreviated) old-style ring (2 node cluster with 2 disks each):

Devices:    id  region  zone      ip address  port  replication ip  replication port      name
             0       1     1       1.1.0.1    6200       1.1.0.1                6200      d1
             1       1     1       1.1.0.1    6200       1.1.0.1                6200      d2
             2       1     2       1.1.0.2    6200       1.1.0.2                6200      d3
             3       1     2       1.1.0.2    6200       1.1.0.2                6200      d4

And here's the same ring set up for servers_per_port:

Devices:    id  region  zone      ip address  port  replication ip  replication port      name
             0       1     1       1.1.0.1    6200       1.1.0.1                6200      d1
             1       1     1       1.1.0.1    6201       1.1.0.1                6201      d2
             2       1     2       1.1.0.2    6200       1.1.0.2                6200      d3
             3       1     2       1.1.0.2    6201       1.1.0.2                6201      d4

When migrating from normal to servers_per_port, perform these steps in order:

1. Upgrade Swift code to a version capable of doing servers_per_port.
2. Enable servers_per_port with a value greater than zero.
3. Restart swift-object-server processes with a SIGHUP. At this point, you will have the servers_per_port number of swift-object-server processes serving all requests for all disks on each node. This preserves availability, but you should perform the next step as quickly as possible.
4. Push out new rings that actually have different ports per disk on each server. One of the ports in the new ring should be the same as the port used in the old ring ("6200" in the example above). This will cover existing proxy-server processes who haven't loaded the new ring yet. They can still talk to any storage node regardless of whether or not that storage node has loaded the ring and started object-server processes on the new ports.

If you do not run a separate object-server for replication, then this setting must be available to the object-replicator and object-reconstructor (i.e. appear in the [DEFAULT] config section).

General Service Configuration

Most Swift services fall into two categories. Swift's wsgi servers and background daemons.

For more information specific to the configuration of Swift's wsgi servers with paste deploy see general-server-configuration.

Configuration for servers and daemons can be expressed together in the same file for each type of server, or separately. If a required section for the service trying to start is missing there will be an error. The sections not used by the service are ignored.

Consider the example of an object storage node. By convention, configuration for the object-server, object-updater, object-replicator, object-auditor, and object-reconstructor exist in a single file /etc/swift/object-server.conf:

[DEFAULT]
reclaim_age = 604800

[pipeline:main]
pipeline = object-server

[app:object-server]
use = egg:swift#object

[object-replicator]

[object-updater]

[object-auditor]

Swift services expect a configuration path as the first argument:

$ swift-object-auditor
Usage: swift-object-auditor CONFIG [options]

Error: missing config path argument

If you omit the object-auditor section this file could not be used as the configuration path when starting the swift-object-auditor daemon:

$ swift-object-auditor /etc/swift/object-server.conf
Unable to find object-auditor config section in /etc/swift/object-server.conf

If the configuration path is a directory instead of a file all of the files in the directory with the file extension ".conf" will be combined to generate the configuration object which is delivered to the Swift service. This is referred to generally as "directory based configuration".

Directory based configuration leverages ConfigParser's native multi-file support. Files ending in ".conf" in the given directory are parsed in lexicographical order. Filenames starting with '.' are ignored. A mixture of file and directory configuration paths is not supported - if the configuration path is a file only that file will be parsed.

The Swift service management tool swift-init has adopted the convention of looking for /etc/swift/{type}-server.conf.d/ if the file /etc/swift/{type}-server.conf file does not exist.

When using directory based configuration, if the same option under the same section appears more than once in different files, the last value parsed is said to override previous occurrences. You can ensure proper override precedence by prefixing the files in the configuration directory with numerical values.:

/etc/swift/
    default.base
    object-server.conf.d/
        000_default.conf -> ../default.base
        001_default-override.conf
        010_server.conf
        020_replicator.conf
        030_updater.conf
        040_auditor.conf

You can inspect the resulting combined configuration object using the swift-config command line tool

General Server Configuration

Swift uses paste.deploy (http://pythonpaste.org/deploy/) to manage server configurations. Detailed descriptions of configuration options can be found in the configuration documentation <config/index>.

Default configuration options are set in the [DEFAULT] section, and any options specified there can be overridden in any of the other sections BUT ONLY BY USING THE SYNTAX set option_name = value. This is the unfortunate way paste.deploy works and I'll try to explain it in full.

First, here's an example paste.deploy configuration file:

[DEFAULT]
name1 = globalvalue
name2 = globalvalue
name3 = globalvalue
set name4 = globalvalue

[pipeline:main]
pipeline = myapp

[app:myapp]
use = egg:mypkg#myapp
name2 = localvalue
set name3 = localvalue
set name5 = localvalue
name6 = localvalue

The resulting configuration that myapp receives is:

global {'__file__': '/etc/mypkg/wsgi.conf', 'here': '/etc/mypkg',
        'name1': 'globalvalue',
        'name2': 'globalvalue',
        'name3': 'localvalue',
        'name4': 'globalvalue',
        'name5': 'localvalue',
        'set name4': 'globalvalue'}
local {'name6': 'localvalue'}

So, name1 got the global value which is fine since it's only in the DEFAULT section anyway.

name2 got the global value from DEFAULT even though it appears to be overridden in the app:myapp subsection. This is just the unfortunate way paste.deploy works (at least at the time of this writing.)

name3 got the local value from the app:myapp subsection because it is using the special paste.deploy syntax of set option_name = value. So, if you want a default value for most app/filters but want to override it in one subsection, this is how you do it.

name4 got the global value from DEFAULT since it's only in that section anyway. But, since we used the set syntax in the DEFAULT section even though we shouldn't, notice we also got a set name4 variable. Weird, but probably not harmful.

name5 got the local value from the app:myapp subsection since it's only there anyway, but notice that it is in the global configuration and not the local configuration. This is because we used the set syntax to set the value. Again, weird, but not harmful since Swift just treats the two sets of configuration values as one set anyway.

name6 got the local value from app:myapp subsection since it's only there, and since we didn't use the set syntax, it's only in the local configuration and not the global one. Though, as indicated above, there is no special distinction with Swift.

That's quite an explanation for something that should be so much simpler, but it might be important to know how paste.deploy interprets configuration files. The main rule to remember when working with Swift configuration files is:

Note

Use the set option_name = value syntax in subsections if the option is also set in the [DEFAULT] section. Don't get in the habit of always using the set syntax or you'll probably mess up your non-paste.deploy configuration files.

Per policy configuration

Some proxy-server configuration options may be overridden for individual overview_policies by including per-policy config section(s). These options are:

• sorting_method
• read_affinity
• write_affinity
• write_affinity_node_count
• write_affinity_handoff_delete_count

The per-policy config section name must be of the form:

[proxy-server:policy:<policy index>]

Note

The per-policy config section name should refer to the policy index, not the policy name.

Note

The first part of proxy-server config section name must match the name of the proxy-server config section. This is typically proxy-server as shown above, but if different then the names of any per-policy config sections must be changed accordingly.

The value of an option specified in a per-policy section will override any value given in the proxy-server section for that policy only. Otherwise the value of these options will be that specified in the proxy-server section.

For example, the following section provides policy-specific options for a policy with index 3:

[proxy-server:policy:3]
sorting_method = affinity
read_affinity = r2=1
write_affinity = r2
write_affinity_node_count = 1 * replicas
write_affinity_handoff_delete_count = 2

Note

It is recommended that per-policy config options are not included in the [DEFAULT] section. If they are then the following behavior applies.

Per-policy config sections will inherit options in the [DEFAULT] section of the config file, and any such inheritance will take precedence over inheriting options from the proxy-server config section.

Per-policy config section options will override options in the [DEFAULT] section. Unlike the behavior described under General Server Configuration for paste-deploy filter and app sections, the set keyword is not required for options to override in per-policy config sections.

For example, given the following settings in a config file:

[DEFAULT]
sorting_method = affinity
read_affinity = r0=100
write_affinity = r0

[app:proxy-server]
use = egg:swift#proxy
# use of set keyword here overrides [DEFAULT] option
set read_affinity = r1=100
# without set keyword, [DEFAULT] option overrides in a paste-deploy section
write_affinity = r1

[proxy-server:policy:0]
sorting_method = affinity
# set keyword not required here to override [DEFAULT] option
write_affinity = r1

would result in policy with index 0 having settings:

• read_affinity = r0=100 (inherited from the [DEFAULT] section)
• write_affinity = r1 (specified in the policy 0 section)

and any other policy would have the default settings of:

• read_affinity = r1=100 (set in the proxy-server section)
• write_affinity = r0 (inherited from the [DEFAULT] section)

Proxy Middlewares

Many features in Swift are implemented as middleware in the proxy-server pipeline. See middleware and the proxy-server.conf-sample file for more information. In particular, the use of some type of authentication and authorization middleware <overview_auth> is highly recommended.

Memcached Considerations

Several of the Services rely on Memcached for caching certain types of lookups, such as auth tokens, and container/account existence. Swift does not do any caching of actual object data. Memcached should be able to run on any servers that have available RAM and CPU. At Rackspace, we run Memcached on the proxy servers. The memcache_servers config option in the proxy-server.conf should contain all memcached servers.

System Time

Time may be relative but it is relatively important for Swift! Swift uses timestamps to determine which is the most recent version of an object. It is very important for the system time on each server in the cluster to by synced as closely as possible (more so for the proxy server, but in general it is a good idea for all the servers). At Rackspace, we use NTP with a local NTP server to ensure that the system times are as close as possible. This should also be monitored to ensure that the times do not vary too much.

General Service Tuning

Most services support either a workers or concurrency value in the settings. This allows the services to make effective use of the cores available. A good starting point to set the concurrency level for the proxy and storage services to 2 times the number of cores available. If more than one service is sharing a server, then some experimentation may be needed to find the best balance.

At Rackspace, our Proxy servers have dual quad core processors, giving us 8 cores. Our testing has shown 16 workers to be a pretty good balance when saturating a 10g network and gives good CPU utilization.

Our Storage server processes all run together on the same servers. These servers have dual quad core processors, for 8 cores total. We run the Account, Container, and Object servers with 8 workers each. Most of the background jobs are run at a concurrency of 1, with the exception of the replicators which are run at a concurrency of 2.

The max_clients parameter can be used to adjust the number of client requests an individual worker accepts for processing. The fewer requests being processed at one time, the less likely a request that consumes the worker's CPU time, or blocks in the OS, will negatively impact other requests. The more requests being processed at one time, the more likely one worker can utilize network and disk capacity.

On systems that have more cores, and more memory, where one can afford to run more workers, raising the number of workers and lowering the maximum number of clients serviced per worker can lessen the impact of CPU intensive or stalled requests.

The nice_priority parameter can be used to set program scheduling priority. The ionice_class and ionice_priority parameters can be used to set I/O scheduling class and priority on the systems that use an I/O scheduler that supports I/O priorities. As at kernel 2.6.17 the only such scheduler is the Completely Fair Queuing (CFQ) I/O scheduler. If you run your Storage servers all together on the same servers, you can slow down the auditors or prioritize object-server I/O via these parameters (but probably do not need to change it on the proxy). It is a new feature and the best practices are still being developed. On some systems it may be required to run the daemons as root. For more info also see setpriority(2) and ioprio_set(2).

The above configuration setting should be taken as suggestions and testing of configuration settings should be done to ensure best utilization of CPU, network connectivity, and disk I/O.

Filesystem Considerations

Swift is designed to be mostly filesystem agnostic--the only requirement being that the filesystem supports extended attributes (xattrs). After thorough testing with our use cases and hardware configurations, XFS was the best all-around choice. If you decide to use a filesystem other than XFS, we highly recommend thorough testing.

For distros with more recent kernels (for example Ubuntu 12.04 Precise), we recommend using the default settings (including the default inode size of 256 bytes) when creating the file system:

mkfs.xfs -L D1 /dev/sda1

In the last couple of years, XFS has made great improvements in how inodes are allocated and used. Using the default inode size no longer has an impact on performance.

For distros with older kernels (for example Ubuntu 10.04 Lucid), some settings can dramatically impact performance. We recommend the following when creating the file system:

mkfs.xfs -i size=1024 -L D1 /dev/sda1

Setting the inode size is important, as XFS stores xattr data in the inode. If the metadata is too large to fit in the inode, a new extent is created, which can cause quite a performance problem. Upping the inode size to 1024 bytes provides enough room to write the default metadata, plus a little headroom.

The following example mount options are recommended when using XFS:

mount -t xfs -o noatime -L D1 /srv/node/d1

We do not recommend running Swift on RAID, but if you are using RAID it is also important to make sure that the proper sunit and swidth settings get set so that XFS can make most efficient use of the RAID array.

For a standard Swift install, all data drives are mounted directly under /srv/node (as can be seen in the above example of mounting label D1 as /srv/node/d1). If you choose to mount the drives in another directory, be sure to set the devices config option in all of the server configs to point to the correct directory.

The mount points for each drive in /srv/node/ should be owned by the root user almost exclusively (root:root 755). This is required to prevent rsync from syncing files into the root drive in the event a drive is unmounted.

Swift uses system calls to reserve space for new objects being written into the system. If your filesystem does not support fallocate() or posix_fallocate(), be sure to set the disable_fallocate = true config parameter in account, container, and object server configs.

Most current Linux distributions ship with a default installation of updatedb. This tool runs periodically and updates the file name database that is used by the GNU locate tool. However, including Swift object and container database files is most likely not required and the periodic update affects the performance quite a bit. To disable the inclusion of these files add the path where Swift stores its data to the setting PRUNEPATHS in /etc/updatedb.conf:

PRUNEPATHS="... /tmp ... /var/spool ... /srv/node"

General System Tuning

Rackspace currently runs Swift on Ubuntu Server 10.04, and the following changes have been found to be useful for our use cases.

The following settings should be in /etc/sysctl.conf:

# disable TIME_WAIT.. wait..
net.ipv4.tcp_tw_recycle=1
net.ipv4.tcp_tw_reuse=1

# disable syn cookies
net.ipv4.tcp_syncookies = 0

# double amount of allowed conntrack
net.ipv4.netfilter.ip_conntrack_max = 262144

To load the updated sysctl settings, run sudo sysctl -p.

A note about changing the TIME_WAIT values. By default the OS will hold a port open for 60 seconds to ensure that any remaining packets can be received. During high usage, and with the number of connections that are created, it is easy to run out of ports. We can change this since we are in control of the network. If you are not in control of the network, or do not expect high loads, then you may not want to adjust those values.

Logging Considerations

Swift is set up to log directly to syslog. Every service can be configured with the log_facility option to set the syslog log facility destination. We recommended using syslog-ng to route the logs to specific log files locally on the server and also to remote log collecting servers. Additionally, custom log handlers can be used via the custom_log_handlers setting.