第 2 章 GFS2 Configuration and Operational Considerations

GFS2 is based on a 64-bit architecture, which can theoretically accommodate an 8 EB file system. However, the current supported maximum size of a GFS2 file system for 64-bit hardware is 100TB.

Note that even though GFS2 large file systems are possible, that does not mean they are recommended. The rule of thumb with GFS2 is that smaller is better: it is better to have 10 1TB file systems than one 10TB file system.

There are several reasons why you should keep your GFS2 file systems small:

In addition, fewer resource groups to maintain mean better performance.

Of course, if you make your GFS2 file system too small, you might run out of space, and that has its own consequences. You should consider your own use cases before deciding on a size.

The mkfs.gfs2 command attempts to estimate an optimal block size based on device topology. In general, 4K blocks are the preferred block size because 4K is the default page size (memory) for Linux. Unlike some other file systems, GFS2 does most of its operations using 4K kernel buffers. If your block size is 4K, the kernel has to do less work to manipulate the buffers.

It is recommended that you use the default block size, which should yield the highest performance. You may need to use a different block size only if you require efficient storage of many very small files.

GFS2 requires one journal for each node in the cluster that needs to mount the file system. For example, if you have a 16-node cluster but need to mount only the file system from two nodes, you need only two journals. If you need to mount from a third node, you can always add a journal with the gfs2_jadd command. With GFS2, you can add journals on the fly.

When you run the mkfs.gfs2 command to create a GFS2 file system, you may specify the size of the journals. If you do not specify a size, it will default to 128MB, which should be optimal for most applications.

Some system administrators might think that 128MB is excessive and be tempted to reduce the size of the journal to the minimum of 8MB or a more conservative 32MB. While that might work, it can severely impact performance. Like many journaling file systems, every time GFS2 writes metadata, the metadata is committed to the journal before it is put into place. This ensures that if the system crashes or loses power, you will recover all of the metadata when the journal is automatically replayed at mount time. However, it does not take much file system activity to fill an 8MB journal, and when the journal is full, performance slows because GFS2 has to wait for writes to the storage.

It is generally recommended to use the default journal size of 128MB. If your file system is very small (for example, 5GB), having a 128MB journal might be impractical. If you have a larger file system and can afford the space, using 256MB journals might improve performance.

When a GFS2 file system is created with the mkfs.gfs2 command, it divides the storage into uniform slices known as resource groups. It attempts to estimate an optimal resource group size (ranging from 32MB to 2GB). You can override the default with the -r option of the mkfs.gfs2 command.

Your optimal resource group size depends on how you will use the file system. Consider how full it will be and whether or not it will be severely fragmented.

You should experiment with different resource group sizes to see which results in optimal performance. It is a best practice to experiment with a test cluster before deploying GFS2 into full production.

If your file system has too many resource groups (each of which is too small), block allocations can waste too much time searching tens of thousands (or hundreds of thousands) of resource groups for a free block. The more full your file system, the more resource groups that will be searched, and every one of them requires a cluster-wide lock. This leads to slow performance.

If, however, your file system has too few resource groups (each of which is too big), block allocations might contend more often for the same resource group lock, which also impacts performance. For example, if you have a 10GB file system that is carved up into five resource groups of 2GB, the nodes in your cluster will fight over those five resource groups more often than if the same file system were carved into 320 resource groups of 32MB. The problem is exacerbated if your file system is nearly full because every block allocation might have to look through several resource groups before it finds one with a free block. GFS2 tries to mitigate this problem in two ways:

The worst case scenario is when there is a central directory in which all the nodes create files because all of the nodes will constantly fight to lock the same resource group.