- Managing the File System and I/O
Sometimes a Lustre file system becomes unbalanced, often due to incorrectly-specified stripe settings, or when very large files are created that are not striped over all of the OSTs. Lustre will automatically avoid allocating new files on OSTs that are full. If an OST is completely full and more data is written to files already located on that OST, an error occurs. The procedures below describe how to handle a full OST.
The MDS will normally handle space balancing automatically at file creation time, and this procedure is normally not needed, but manual data migration may be desirable in some cases (e.g. creating very large files that would consume more than the total free space of the full OSTs).
Checking OST Space Usage
The example below shows an unbalanced file system:
client# lfs df -h
UUID bytes Used Available \
Use% Mounted on
testfs-MDT0000_UUID 4.4G 214.5M 3.9G \
4% /mnt/testfs[MDT:0]
testfs-OST0000_UUID 2.0G 751.3M 1.1G \
37% /mnt/testfs[OST:0]
testfs-OST0001_UUID 2.0G 755.3M 1.1G \
37% /mnt/testfs[OST:1]
testfs-OST0002_UUID 2.0G 1.7G 155.1M \
86% /mnt/testfs[OST:2] ****
testfs-OST0003_UUID 2.0G 751.3M 1.1G \
37% /mnt/testfs[OST:3]
testfs-OST0004_UUID 2.0G 747.3M 1.1G \
37% /mnt/testfs[OST:4]
testfs-OST0005_UUID 2.0G 743.3M 1.1G \
36% /mnt/testfs[OST:5]
filesystem summary: 11.8G 5.4G 5.8G \
45% /mnt/testfs
In this case, OST0002 is almost full and when an attempt is made to write additional information to the file system (even with uniform striping over all the OSTs), the write command fails as follows:
client# lfs setstripe /mnt/testfs 4M 0 -1
client# dd if=/dev/zero of=/mnt/testfs/test_3 bs=10M count=100
dd: writing '/mnt/testfs/test_3': No space left on device
98+0 records in
97+0 records out
1017192448 bytes (1.0 GB) copied, 23.2411 seconds, 43.8 MB/s
The example below shows an unbalanced file system:
client# lfs df -h
UUID bytes Used Available \
Use% Mounted on
testfs-MDT0000_UUID 4.4G 214.5M 3.9G \
4% /mnt/testfs[MDT:0]
testfs-OST0000_UUID 2.0G 751.3M 1.1G \
37% /mnt/testfs[OST:0]
testfs-OST0001_UUID 2.0G 755.3M 1.1G \
37% /mnt/testfs[OST:1]
testfs-OST0002_UUID 2.0G 1.7G 155.1M \
86% /mnt/testfs[OST:2] ****
testfs-OST0003_UUID 2.0G 751.3M 1.1G \
37% /mnt/testfs[OST:3]
testfs-OST0004_UUID 2.0G 747.3M 1.1G \
37% /mnt/testfs[OST:4]
testfs-OST0005_UUID 2.0G 743.3M 1.1G \
36% /mnt/testfs[OST:5]
filesystem summary: 11.8G 5.4G 5.8G \
45% /mnt/testfs
In this case, OST0002 is almost full and when an attempt is made to write additional information to the file system (even with uniform striping over all the OSTs), the write command fails as follows:
client# lfs setstripe /mnt/testfs 4M 0 -1
client# dd if=/dev/zero of=/mnt/testfs/test_3 bs=10M count=100
dd: writing '/mnt/testfs/test_3': No space left on device
98+0 records in
97+0 records out
1017192448 bytes (1.0 GB) copied, 23.2411 seconds, 43.8 MB/s
To avoid running out of space in the file system, if the OST usage is imbalanced and one or more OSTs are close to being full while there are others that have a lot of space, the MDS will typically avoid file creation on the full OST(s) automatically. The full OSTs may optionally be deactivated manually on the MDS to ensure the MDS will not allocate new objects there.
-
Log into the MDS server and use the
lctlcommand to stop new object creation on the full OST(s):mds# lctl set_param osp.fsname-OSTnnnn*.max_create_count=0
When new files are created in the file system, they will only use the remaining OSTs. Either manual space rebalancing can be done by migrating data to other OSTs, as shown in the next section, or normal file deletion and creation can passively rebalance the space usage.
If there is a need to move the file data from the current OST(s) to new OST(s), the data must be migrated (copied) to the new location. The simplest way to do this is to use the lfs_migrate command, as described in the section called “ Adding a New OST to a Lustre File System”.
Once the full OST(s) no longer are severely imbalanced, due to either active or passive data redistribution, they should be reactivated so they will again have new files allocated on them.
[mds]# lctl set_param osp.testfs-OST0002.max_create_count=20000
Introduced in Lustre 2.8
Lustre software version 2.8 includes a feature to migrate metadata (directories and inodes therein) between MDTs. This migration can only be performed on whole directories. Striped directories are not supported until Lustre 2.12. For example, to migrate the contents of the /testfs/remotedir directory from the MDT where it currently is located to MDT0000 to allow that MDT to be removed, the sequence of commands is as follows:
$ cd /testfs
$ lfs getdirstripe -m ./remotedir *which MDT is dir on?*
1
$ touch ./remotedir/file.{1,2,3}.txt*create test files*
$ lfs getstripe -m ./remotedir/file.*.txt*check files are on MDT0001*
1
1
1
$ lfs migrate -m 0 ./remotedir *migrate testremote to MDT0000*
$ lfs getdirstripe -m ./remotedir *which MDT is dir on now?*
0
$ lfs getstripe -m ./remotedir/file.*.txt*check files are on MDT0000*
0
0
0
For more information, see man lfs-migrate.
During migration each file receives a new identifier (FID). As a consequence, the file will report a new inode number to userspace applications. Some system tools (for example, backup and archiving tools, NFS, Samba) that identify files by inode number may consider the migrated files to be new, even though the contents are unchanged. If a Lustre system is re-exporting to NFS, the migrated files may become inaccessible during and after migration if the client or server are caching a stale file handle with the old FID. Restarting the NFS service will flush the local file handle cache, but clients may also need to be restarted as they may cache stale file handles as well.
Introduced in Lustre 2.12
Lustre 2.8 included a feature to migrate metadata (directories and inodes therein) between MDTs, however it did not support migration of striped directories, or changing the stripe count of an existing directory. Lustre 2.12 adds support for migrating and restriping directories. The lfs migrate -m command can only only be performed on whole directories, though it will migrate both the specified directory and its sub-entries recursively. For example, to migrate the contents of a large directory /testfs/largedir from its current location on MDT0000 to MDT0001 and MDT0003, run the following command:
$ lfs migrate -m 1,3 /testfs/largedir
Metadata migration will migrate file dirent and inode to other MDTs, but it won't touch file data. During migration, directory and its sub-files can be accessed like normal ones, though the same warning above applies to tools that depend on the file inode number. Migration may fail for various reasons such as MDS restart, or disk full. In those cases, some of the sub-files may have been migrated to the new MDTs, while others are still on the original MDT. The files can be accessed normally. The same lfs migrate -m command should be executed again when these issues are fixed to finish this migration. However, you cannot abort a failed migration, or migrate to different MDTs from previous migration command.
The OST pools feature enables users to group OSTs together to make object placement more flexible. A 'pool' is the name associated with an arbitrary subset of OSTs in a Lustre cluster.
OST pools follow these rules:
- An OST can be a member of multiple pools.
- No ordering of OSTs in a pool is defined or implied.
- Stripe allocation within a pool follows the same rules as the normal stripe allocator.
- OST membership in a pool is flexible, and can change over time.
When an OST pool is defined, it can be used to allocate files. When file or directory striping is set to a pool, only OSTs in the pool are candidates for striping. If a stripe_index is specified which refers to an OST that is not a member of the pool, an error is returned.
OST pools are used only at file creation. If the definition of a pool changes (an OST is added or removed or the pool is destroyed), already-created files are not affected.
Note
An error ( EINVAL) results if you create a file using an empty pool.
Note
If a directory has pool striping set and the pool is subsequently removed, the new files created in this directory have the (non-pool) default striping pattern for that directory applied and no error is returned.
OST pools are defined in the configuration log on the MGS. Use the lctl command to:
- Create/destroy a pool
- Add/remove OSTs in a pool
- List pools and OSTs in a specific pool
The lctl command MUST be run on the MGS. Another requirement for managing OST pools is to either have the MDT and MGS on the same node or have a Lustre client mounted on the MGS node, if it is separate from the MDS. This is needed to validate the pool commands being run are correct.
Caution
Running the writeconf command on the MDS erases all pools information (as well as any other parameters set using lctl conf_param). We recommend that the pools definitions (and conf_paramsettings) be executed using a script, so they can be reproduced easily after a writeconf is performed.
To create a new pool, run:
mgs# lctl pool_new
fsname.
poolname
Note
The pool name is an ASCII string up to 15 characters.
To add the named OST to a pool, run:
mgs# lctl pool_add
fsname.
poolname
ost_list
Where:
*ost_list*is *fsname*-OST *index_range**index_range*is *ost_index_start*- *ost_index_end[,index_range]*or*ost_index_start*-*ost_index_end/step*
If the leading *fsname* and/or ending _UUID are missing, they are automatically added.
For example, to add even-numbered OSTs to pool1 on file system testfs, run a single command ( pool_add) to add many OSTs to the pool at one time:
lctl pool_add testfs.pool1 OST[0-10/2]
Note
Each time an OST is added to a pool, a new llog configuration record is created. For convenience, you can run a single command.
To remove a named OST from a pool, run:
mgs# lctl pool_remove
fsname.
poolname
ost_list
To destroy a pool, run:
mgs# lctl pool_destroy
fsname.
poolname
Note
All OSTs must be removed from a pool before it can be destroyed.
To list pools in the named file system, run:
mgs# lctl pool_list
fsname|pathname
To list OSTs in a named pool, run:
lctl pool_list
fsname.
poolname
Several lfs commands can be run with OST pools. Use the lfs setstripe command to associate a directory with an OST pool. This causes all new regular files and directories in the directory to be created in the pool. The lfs command can be used to list pools in a file system and OSTs in a named pool.
To associate a directory with a pool, so all new files and directories will be created in the pool, run:
client# lfs setstripe --pool|-p pool_name
filename|dirname
To set striping patterns, run:
client# lfs setstripe [--size|-s stripe_size] [--offset|-o start_ost]
[--count|-c stripe_count] [--pool|-p pool_name]
dir|filename
**Note **
If you specify striping with an invalid pool name, because the pool does not exist or the pool name was mistyped, lfs setstripe returns an error. Run lfs pool_list to make sure the pool exists and the pool name is entered correctly.
Note
The --pool option for lfs setstripe is compatible with other modifiers. For example, you can set striping on a directory to use an explicit starting index.
Here are several suggestions for using OST pools.
- A directory or file can be given an extended attribute (EA), that restricts striping to a pool.
- Pools can be used to group OSTs with the same technology or performance (slower or faster), or that are preferred for certain jobs. Examples are SATA OSTs versus SAS OSTs or remote OSTs versus local OSTs.
- A file created in an OST pool tracks the pool by keeping the pool name in the file LOV EA.
To add an OST to existing Lustre file system:
-
Add a new OST by passing on the following commands, run:
oss# mkfs.lustre --fsname=testfs --mgsnode=mds16@tcp0 --ost --index=12 /dev/sda oss# mkdir -p /mnt/testfs/ost12 oss# mount -t lustre /dev/sda /mnt/testfs/ost12 -
Migrate the data (possibly).
The file system is quite unbalanced when new empty OSTs are added. New file creations are automatically balanced. If this is a scratch file system or files are pruned at a regular interval, then no further work may be needed. Files existing prior to the expansion can be rebalanced with an in-place copy, which can be done with a simple script.
The basic method is to copy existing files to a temporary file, then move the temp file over the old one. This should not be attempted with files which are currently being written to by users or applications. This operation redistributes the stripes over the entire set of OSTs.
A very clever migration script would do the following:
- Examine the current distribution of data.
- Calculate how much data should move from each full OST to the empty ones.
- Search for files on a given full OST (using
lfs getstripe). - Force the new destination OST (using
lfs setstripe). - Copy only enough files to address the imbalance.
If a Lustre file system administrator wants to explore this approach further, per-OST disk-usage statistics can be found under /proc/fs/lustre/osc/*/rpc_stats
The Lustre software supports the O_DIRECT flag to open.
Applications using the read() and write() calls must supply buffers aligned on a page boundary (usually 4 K). If the alignment is not correct, the call returns -EINVAL. Direct I/O may help performance in cases where the client is doing a large amount of I/O and is CPU-bound (CPU utilization 100%).
An immutable file or directory is one that cannot be modified, renamed or removed. To do this:
chattr +i
file
To remove this flag, use chattr -i
This section describes other I/O options, including checksums, and the ptlrpcd thread pool.
To guard against network data corruption, a Lustre client can perform two types of data checksums: in-memory (for data in client memory) and wire (for data sent over the network). For each checksum type, a 32-bit checksum of the data read or written on both the client and server is computed, to ensure that the data has not been corrupted in transit over the network. The ldiskfs backing file system does NOT do any persistent checksumming, so it does not detect corruption of data in the OST file system.
The checksumming feature is enabled, by default, on individual client nodes. If the client or OST detects a checksum mismatch, then an error is logged in the syslog of the form:
LustreError: BAD WRITE CHECKSUM: changed in transit before arrival at OST: \
from 192.168.1.1@tcp inum 8991479/2386814769 object 1127239/0 extent [10240\
0-106495]
If this happens, the client will re-read or re-write the affected data up to five times to get a good copy of the data over the network. If it is still not possible, then an I/O error is returned to the application.
To enable both types of checksums (in-memory and wire), run:
lctl set_param llite.*.checksum_pages=1
To disable both types of checksums (in-memory and wire), run:
lctl set_param llite.*.checksum_pages=0
To check the status of a wire checksum, run:
lctl get_param osc.*.checksums
By default, the Lustre software uses the adler32 checksum algorithm, because it is robust and has a lower impact on performance than crc32. The Lustre file system administrator can change the checksum algorithm via lctl get_param, depending on what is supported in the kernel.
To check which checksum algorithm is being used by the Lustre software, run:
$ lctl get_param osc.*.checksum_type
To change the wire checksum algorithm, run:
$ lctl set_param osc.*.checksum_type=
algorithm
**Note **
The in-memory checksum always uses the adler32 algorithm, if available, and only falls back to crc32 if adler32 cannot be used.
In the following example, the lctl get_param command is used to determine that the Lustre software is using the adler32 checksum algorithm. Then the lctl set_param command is used to change the checksum algorithm to crc32. A second lctl get_param command confirms that the crc32 checksum algorithm is now in use.
$ lctl get_param osc.*.checksum_type
osc.testfs-OST0000-osc-ffff81012b2c48e0.checksum_type=crc32 [adler]
$ lctl set_param osc.*.checksum_type=crc32
osc.testfs-OST0000-osc-ffff81012b2c48e0.checksum_type=crc32
$ lctl get_param osc.*.checksum_type
osc.testfs-OST0000-osc-ffff81012b2c48e0.checksum_type=[crc32] adler
Releases prior to Lustre software release 2.2 used two portal RPC daemons for each client/server pair. One daemon handled all synchronous IO requests, and the second daemon handled all asynchronous (non-IO) RPCs. The increasing use of large SMP nodes for Lustre servers exposed some scaling issues. The lack of threads for large SMP nodes resulted in cases where a single CPU would be 100% utilized and other CPUs would be relativity idle. This is especially noticeable when a single client traverses a large directory.
Lustre software release 2.2.x implements a ptlrpc thread pool, so that multiple threads can be created to serve asynchronous RPC requests. The number of threads spawned is controlled at module load time using module options. By default one thread is spawned per CPU, with a minimum of 2 threads spawned irrespective of module options.
One of the issues with thread operations is the cost of moving a thread context from one CPU to another with the resulting loss of CPU cache warmth. To reduce this cost, ptlrpc threads can be bound to a CPU. However, if the CPUs are busy, a bound thread may not be able to respond quickly, as the bound CPU may be busy with other tasks and the thread must wait to schedule.
Because of these considerations, the pool of ptlrpc threads can be a mixture of bound and unbound threads. The system operator can balance the thread mixture based on system size and workload.
These parameters should be set in /etc/modprobe.conf or in the etc/modprobe.d directory, as options for the ptlrpc module.
options ptlrpcd max_ptlrpcds=XXX
Sets the number of ptlrpcd threads created at module load time. The default if not specified is one thread per CPU, including hyper-threaded CPUs. The lower bound is 2 (old prlrpcd behaviour)
options ptlrpcd ptlrpcd_bind_policy=[1-4]
Controls the binding of threads to CPUs. There are four policy options.
PDB_POLICY_NONE(ptlrpcd_bind_policy=1) All threads are unbound.PDB_POLICY_FULL(ptlrpcd_bind_policy=2) All threads attempt to bind to a CPU.PDB_POLICY_PAIR(ptlrpcd_bind_policy=3) This is the default policy. Threads are allocated as a bound/unbound pair. Each thread (bound or free) has a partner thread. The partnering is used by the ptlrpcd load policy, which determines how threads are allocated to CPUs.PDB_POLICY_NEIGHBOR(ptlrpcd_bind_policy=4) Threads are allocated as a bound/unbound pair. Each thread (bound or free) has two partner threads.