qcow2 L2/refcount cache configuration ===================================== Copyright (C) 2015, 2018 Igalia, S.L. Author: Alberto Garcia This work is licensed under the terms of the GNU GPL, version 2 or later. See the COPYING file in the top-level directory. Introduction ------------ The QEMU qcow2 driver has two caches that can improve the I/O performance significantly. However, setting the right cache sizes is not a straightforward operation. This document attempts to give an overview of the L2 and refcount caches, and how to configure them. Please refer to the docs/interop/qcow2.txt file for an in-depth technical description of the qcow2 file format. Clusters -------- A qcow2 file is organized in units of constant size called clusters. The cluster size is configurable, but it must be a power of two and its value 512 bytes or higher. QEMU currently defaults to 64 KB clusters, and it does not support sizes larger than 2MB. The 'qemu-img create' command supports specifying the size using the cluster_size option: qemu-img create -f qcow2 -o cluster_size=128K hd.qcow2 4G The L2 tables ------------- The qcow2 format uses a two-level structure to map the virtual disk as seen by the guest to the disk image in the host. These structures are called the L1 and L2 tables. There is one single L1 table per disk image. The table is small and is always kept in memory. There can be many L2 tables, depending on how much space has been allocated in the image. Each table is one cluster in size. In order to read or write data from the virtual disk, QEMU needs to read its corresponding L2 table to find out where that data is located. Since reading the table for each I/O operation can be expensive, QEMU keeps an L2 cache in memory to speed up disk access. The size of the L2 cache can be configured, and setting the right value can improve the I/O performance significantly. The refcount blocks ------------------- The qcow2 format also mantains a reference count for each cluster. Reference counts are used for cluster allocation and internal snapshots. The data is stored in a two-level structure similar to the L1/L2 tables described above. The second level structures are called refcount blocks, are also one cluster in size and the number is also variable and dependent on the amount of allocated space. Each block contains a number of refcount entries. Their size (in bits) is a power of two and must not be higher than 64. It defaults to 16 bits, but a different value can be set using the refcount_bits option: qemu-img create -f qcow2 -o refcount_bits=8 hd.qcow2 4G QEMU keeps a refcount cache to speed up I/O much like the aforementioned L2 cache, and its size can also be configured. Choosing the right cache sizes ------------------------------ In order to choose the cache sizes we need to know how they relate to the amount of allocated space. The part of the virtual disk that can be mapped by the L2 and refcount caches (in bytes) is: disk_size = l2_cache_size * cluster_size / 8 disk_size = refcount_cache_size * cluster_size * 8 / refcount_bits With the default values for cluster_size (64KB) and refcount_bits (16), this becomes: disk_size = l2_cache_size * 8192 disk_size = refcount_cache_size * 32768 So in order to cover n GB of disk space with the default values we need: l2_cache_size = disk_size_GB * 131072 refcount_cache_size = disk_size_GB * 32768 For example, 1MB of L2 cache is needed to cover every 8 GB of the virtual image size (given that the default cluster size is used): 8 GB / 8192 = 1 MB The refcount cache is 4 times the cluster size by default. With the default cluster size of 64 KB, it is 256 KB (262144 bytes). This is sufficient for 8 GB of image size: 262144 * 32768 = 8 GB How to configure the cache sizes -------------------------------- Cache sizes can be configured using the -drive option in the command-line, or the 'blockdev-add' QMP command. There are three options available, and all of them take bytes: "l2-cache-size": maximum size of the L2 table cache "refcount-cache-size": maximum size of the refcount block cache "cache-size": maximum size of both caches combined There are a few things that need to be taken into account: - Both caches must have a size that is a multiple of the cluster size (or the cache entry size: see "Using smaller cache sizes" below). - The maximum L2 cache size is 32 MB by default on Linux platforms (enough for full coverage of 256 GB images, with the default cluster size). This value can be modified using the "l2-cache-size" option. QEMU will not use more memory than needed to hold all of the image's L2 tables, regardless of this max. value. On non-Linux platforms the maximal value is smaller by default (8 MB) and this difference stems from the fact that on Linux the cache can be cleared periodically if needed, using the "cache-clean-interval" option (see below). The minimal L2 cache size is 2 clusters (or 2 cache entries, see below). - The default (and minimum) refcount cache size is 4 clusters. - If only "cache-size" is specified then QEMU will assign as much memory as possible to the L2 cache before increasing the refcount cache size. - At most two of "l2-cache-size", "refcount-cache-size", and "cache-size" can be set simultaneously. Unlike L2 tables, refcount blocks are not used during normal I/O but only during allocations and internal snapshots. In most cases they are accessed sequentially (even during random guest I/O) so increasing the refcount cache size won't have any measurable effect in performance (this can change if you are using internal snapshots, so you may want to think about increasing the cache size if you use them heavily). Before QEMU 2.12 the refcount cache had a default size of 1/4 of the L2 cache size. This resulted in unnecessarily large caches, so now the refcount cache is as small as possible unless overridden by the user. Using smaller cache entries --------------------------- The qcow2 L2 cache stores complete tables by default. This means that if QEMU needs an entry from an L2 table then the whole table is read from disk and is kept in the cache. If the cache is full then a complete table needs to be evicted first. This can be inefficient with large cluster sizes since it results in more disk I/O and wastes more cache memory. Since QEMU 2.12 you can change the size of the L2 cache entry and make it smaller than the cluster size. This can be configured using the "l2-cache-entry-size" parameter: -drive file=hd.qcow2,l2-cache-size=2097152,l2-cache-entry-size=4096 Some things to take into account: - The L2 cache entry size has the same restrictions as the cluster size (power of two, at least 512 bytes). - Smaller entry sizes generally improve the cache efficiency and make disk I/O faster. This is particularly true with solid state drives so it's a good idea to reduce the entry size in those cases. With rotating hard drives the situation is a bit more complicated so you should test it first and stay with the default size if unsure. - Try different entry sizes to see which one gives faster performance in your case. The block size of the host filesystem is generally a good default (usually 4096 bytes in the case of ext4). - Only the L2 cache can be configured this way. The refcount cache always uses the cluster size as the entry size. - If the L2 cache is big enough to hold all of the image's L2 tables (as explained in the "Choosing the right cache sizes" and "How to configure the cache sizes" sections in this document) then none of this is necessary and you can omit the "l2-cache-entry-size" parameter altogether. Reducing the memory usage ------------------------- It is possible to clean unused cache entries in order to reduce the memory usage during periods of low I/O activity. The parameter "cache-clean-interval" defines an interval (in seconds). All cache entries that haven't been accessed during that interval are removed from memory. This example removes all unused cache entries every 15 minutes: -drive file=hd.qcow2,cache-clean-interval=900 If unset, the default value for this parameter is 600. Setting it to 0 disables this feature. Note that this functionality currently relies on the MADV_DONTNEED argument for madvise() to actually free the memory. This is a Linux-specific feature, so cache-clean-interval is not supported in other systems.