LJ Archive

Advanced Hard Drive Caching Techniques

Petros Koutoupis

Issue #233, September 2013

Enhance the longevity of your slower mechanical hard drive and limit its power consumption by caching data to SSD or RAM.

With the introduction of the solid-state Flash drive, performance came to the forefront for data storage technologies. Prior to that, software developers and server administrators needed to devise methods for which they could increase I/O throughput to storage, most of which resulted in low capacity caching to random access memory (RAM) or a RAM drive. Although not as fast as RAM, the Flash drive was almost a dream come true, but it had its limitations—one of which was its low capacities packaged in the NAND-based chips. The traditional spinning disk drive provided users' desired capacities but lacked in speedy accessibility. Even with the 6Gb SATA protocol, sequential data access at best performed at approximately 150MB per second (or MB/s) for both read and write operations, while random access varied between 2–5MB/s as the seeking across multiple sectors laid out in multiple tracks across multiple spinning platters proved to be an extremely disruptive bottleneck. The solid-state drive (SSD) with no movable components significantly decreased these access latencies, thus rendering this bottleneck almost nonexistent.

Even today, the consumer SSD cannot compare to the capacities provided by the magnetic hard disk drive (or HDD), which is why in this article I intend to introduce readers to proven methods for obtaining near SSD performance with the traditional HDD. Multiple open-source projects exist that can achieve this, all but one of which utilizes an SSD as a caching node, and the other caches to RAM. The device drivers I cover here are dm-cache, FlashCache and the RapidDisk/RapidCache suite; I also briefly discuss bcache and EnhanceIO.

In my examples, I am using a commercial SATA III (6Gbps) SSD with an average performance of the following:

  • Sequential read: 231MB/s

  • Sequential write: 74MB/s

  • Random read: 230MB/s

  • Random write: 72MB/s

This SSD provides the caching layer for a slower mechanical SATA III HDD that performs at the following:

  • Sequential read: 115MB/s

  • Sequential write: 72MB/s

  • Random read: 2MB/s

  • Random write: 2MB/s

In my environment, the SSD is labeled as /dev/sdb, and the HDD is /dev/sda3. These are non-intrusive transparent caching solutions intended to achieve the performance benefits of SSDs. They can be added and removed to existing storage targets without issue or data loss (assuming that all cached data has been flushed to disk successfully). Also, all the examples here showcase a write-back caching scheme with the exception of RapidCache, which instead will be used in write-through mode. In write-back mode, newly written data is cached but not immediately written to the destination target. Write-through mode always will write new data to the target while still maintaining it in cache for future reads.

dm-cache

dm-cache has been around for quite some time—at least since 2006. It originally made its debut as a research project developed by Dr Ming Zhao through his summer internship at IBM research. The dm-cache module just recently was integrated into the Linux kernel tree as of version 3.9. Whether you choose to enable it in a recently downloaded kernel or compile it from the official project site, the results will be the same. To load the module, you need to invoke modprobe or insmod:

$ sudo modprobe dm-cache

Now that the module is loaded, you need to inform that module about which drive to point to for the cache and which to point to for the destination. The dm-cache project site provides a Perl script to simplify this process called dmc-setup.pl. For example, if I wanted to use the entire SSD in write-back caching mode with a 4KB block size, I would type:

$ sudo perl dmc-setup.pl -o /dev/sda3 -c /dev/sdb -n cache -b 8 -w

This script is a wrapper to the equivalent dmsetup command below:

$ echo 0 20971520 cache /dev/sda3 /dev/sdb 0 8 65536 16 1 | 
 ↪dmsetup create cache

The dm-cache documentation hosted on the project site provides details on each parameter field, so I don't cover them here.

You may notice that in both examples, I named the mapping to both drives “cache”. So, when I need to access the drive mapping, I must refer to it as “cache”.

The following mapping passes all data requests to the caching driver, which in turn performs the necessary magic to process the requests either by handling it entirely out of cache or both the cache and the slower device:

$ ls -l /dev/mapper
total 0
lrwxrwxrwx 1 root root       7 Jun 30 12:10 cache -> ../dm-0
crw------- 1 root root 10, 236 Jun 30 11:52 control

Just like with any other device-mapper-enabled target, I also can pull up detailed mapping data:

$ sudo dmsetup status cache
0 20971520 cache stats: reads(83), writes(0), 
 ↪cache hits(0, 0.0),replacement(0), replaced dirty blocks(0)

$ sudo dmsetup table cache
0 20971520 cache conf: capacity(256M), associativity(16), 
 ↪block size(4K), write-back

If the target drive already is formatted with data on it, you just need to mount it; otherwise, format it to your specified filesystem:

$ sudo mke2fs -F /dev/mapper/cache

Remember, these solutions are non-intrusive, so if you have existing data that needs to remain on that disk drive, skip the above step and go straight to mounting it for data accessibility:

$ sudo mount /dev/mapper/cache /mnt/cache
$ df|grep cache
/dev/mapper/cache  10321208 1072632   8724288  11% /mnt/cache

Using a benchmarking utility, the numbers will vary. On read operations, it is wholly dependent on whether the desired data resides in cache or whether the module needs to retrieve it from the slower disk. On write operations, it depends on the Flash technology itself, and whether it needs to go through a typical programmable erase (PE) cycle to write the new data. Regardless of this, the random read/write access to the slower drive has been increased significantly:

  • Sequential read: 105MB/s

  • Sequential write: 50MB/s

  • Random read: 67MB/s

  • Random write: 51MB/s

You can continue monitoring the cache status by typing:

$ sudo dmsetup status cache 
0 20971520 cache stats: reads(301319), writes(353216), 
 ↪cache hits(24485, 0.3),replacement(345972), 
 ↪replaced dirty blocks(92857)

To remove the cache mapping, unmount the drive and invoke dmsetup:

$ sudo umount /mnt/cache
$ sudo dmsetup remove cache

FlashCache

FlashCache is a project developed and maintained by Facebook. It was inspired by dm-cache. Much like dm-cache, it too is built from the device-mapper framework. It currently is hosted on GitHub and can be cloned from there. The repository encompasses the kernel module and administration utilities. Once built and installed, load the kernel module and in a similar fashion to the previous examples, create a mapping of the SSD and HDD:

$ sudo modprobe flashcache
$ sudo flashcache_create -p back -b 8 cache /dev/sdb /dev/sda3
cachedev cache, ssd_devname /dev/sdb, disk_devname /dev/sda3 
 ↪cache mode WRITE_BACK block_size 8, md_block_size 8, 
 ↪cache_size 0
FlashCache metadata will use 223MB of your 3944MB main memory

The flashcache_create administration utility is similar to the dmc-setup.pl Perl script used for dm-cache. It is a wrapper utility designed to simplify the dmsetup process. As with the dm-cache module, once the mapping has been created, you can view mapping details by typing:

$ sudo dmsetup table cache
0 20971520 flashcache conf:
    ssd dev (/dev/sdb), disk dev (/dev/sda3) cache mode(WRITE_BACK)
    capacity(57018M), associativity(512), data block size(4K) 
     ↪metadata block size(4096b)
    skip sequential thresh(0K)
    total blocks(14596608), cached blocks(83), cache percent(0)
    dirty blocks(0), dirty percent(0)
    nr_queued(0)
Size Hist: 4096:83 
$ sudo dmsetup status cache
0 20971520 flashcache stats: 
    reads(83), writes(0)
    read hits(0), read hit percent(0)
    write hits(0) write hit percent(0)
    dirty write hits(0) dirty write hit percent(0)
    replacement(0), write replacement(0)
    write invalidates(0), read invalidates(0)
    pending enqueues(0), pending inval(0)
    metadata dirties(0), metadata cleans(0)
    metadata batch(0) metadata ssd writes(0)
    cleanings(0) fallow cleanings(0)
    no room(0) front merge(0) back merge(0)
    disk reads(83), disk writes(0) ssd reads(0) ssd writes(83)
    uncached reads(0), uncached writes(0), uncached IO requeue(0)
    disk read errors(0), disk write errors(0) ssd read errors(0) 
     ↪ssd write errors(0)
    uncached sequential reads(0), uncached sequential writes(0)
    pid_adds(0), pid_dels(0), pid_drops(0) pid_expiry(0)

Mount the mapping for file accessibility:

$ sudo mount /dev/mapper/cache /mnt/cache

Using the same benchmarking utility, observe the differences between FlashCache and the previous module:

  • Sequential read: 284MB/s

  • Sequential write: 72MB/s

  • Random read: 284MB/s

  • Random write: 71MB/s

The numbers look more like the native SSD performance. However, I want to note that this article is not intended to prove that one solution performs better than the other, but instead to enlighten readers of the many methods you can use to accelerate data access to existing and slower configurations.

To unmount and remove the drive mapping, type the following in the terminal:

$ sudo umount /mnt/cache
$ sudo dmsetup remove /dev/mapper/cache

RapidDisk and RapidCache

Currently at version 2.9, RapidDisk is an advanced Linux RAM disk whose features include the capabilities to allocate RAM dynamically as a block device, use it as standalone disk drives, or even map it as caching nodes to slower local disk drives via RapidCache (the latter of which was inspired by FlashCache and uses the device-mapper framework). RAM is being accessed to handle the data storage by allocating memory pages as they are needed. It is a volatile form of storage, so if power is removed or if the computer is rebooted, all data stored within RAM will not be preserved. This is why the RapidCache module was designed to handle only read-through/write-through caching, which means that whatever is intended to be written to the slower storage device will be cached to RapidCache and written immediately to the hard drive. And, if data is being requested from the hard drive and it does not pre-exist in the RapidCache node, it will read the data from the slower device and then cache it to the RapidCache node. This method will retain the same write performance as the hard drive, but significantly increase sequential and random access read performance to cached data.

Once the package, which consists of two kernel modules and an administration utility, is built and installed, you need to insert the modules by typing the following on the command line:

$ sudo modprobe rxdsk
$ sudo modprobe -r rxdsk

Let's assume that you're running on a computer that contains 4GB of RAM, and you confidently can say that at least 1GB of that RAM is never used by the operating system and its applications. Using RapidDisk to create a RAM drive of 1GB in size, you would type:

$ sudo rxadm --attach 1024

Remember, RapidDisk will not pre-allocate this storage. It will allocate RAM only as it is used.

A quick benchmark test of just the RAM drive produces some overwhelmingly fast results with 4KB I/O transfers:

  • Sequential read: 1.6GB/s

  • Sequential write: 1.6GB/s

  • Random read: 1.3GB/s

  • Random write: 1.1GB/s

It produces the following with 1MB I/O transfers:

  • Sequential read: 4.9GB/s

  • Sequential write: 4.3GB/s

  • Random read: 4.9GB/s

  • Random write: 4.0GB/s

Impressive, right? To utilize such a speedy RAM drive as a caching node to a slower drive, a mapping must be created, where /dev/rxd0 is the node used to access the RAM drive, and /dev/mapper/rxc0 is the node used to access the mapping of the two drives:

$ sudo rxadm --rxc-map rxd0 /dev/sda3 4

You can get a list of attached devices and mappings by typing:

$ sudo rxadm --list
rxadm 2.9
Copyright 2011-2013 Petros Koutoupis

List of rxdsk device(s):

 RapidDisk Device 1: rxd0
    Size: 1073741824

List of rxcache mapping(s):

 RapidCache Target 1: rxc0
0 20971519 rxcache conf:
    rxd dev (/dev/rxd0), disk dev (/dev/sda3) mode (WRITETHROUGH)
    capacity(1024M), associativity(512), block size(4K)
    total blocks(262144), cached blocks(0)
 Size Hist: 512:663 

As with the previous device-mapper-based solutions, you even can list detailed information of the mapping by typing:

$ sudo dmsetup table rxc0
0 20971519 rxcache conf:
    rxd dev (/dev/rxd0), disk dev (/dev/sda3) mode (WRITETHROUGH)
    capacity(1024M), associativity(512), block size(4K)
    total blocks(262144), cached blocks(0)
 Size Hist: 512:663 

$ sudo dmsetup status rxc0
0 20971519 rxcache stats: 
    reads(663), writes(0)
    cache hits(0) replacement(0), write replacement(0)
    read invalidates(0), write invalidates(0)
    uncached reads(663), uncached writes(0)
    disk reads(663), disk writes(0)
    cache reads(0), cache writes(0)

Format the mapping if needed and mount it:

$ sudo mount /dev/mapper/rxc0 /mnt/cache

A benchmark test produces the following results:

  • Sequential read: 794MB/s

  • Sequential write: 70MB/s

  • Random read: 901MB/s

  • Random write: 2MB/s

Notice that the write performance is not very great, and that's because it is not meant to be. Write-through mode promises only faster read performance of cached data and consistent write performance to the original drive. The read performance, however, shows significant improvement when accessing cached data.

To remove the mapping and detach the RAM drive, type the following:

$ sudo umount /mnt/cache
$ sudo rxadm --rxc-unmap rxc0
$ sudo rxadm --detach rxd0

Other Solutions Worth Mentioning

bcache:

bcache is relatively new to the hard drive caching scene. It offers all the same features and functionalities as the previous solutions with the exception of its capability to map one or more SSDs as the cache for one or more HDDs instead of one volume to one volume. The project's maintainer does, however, tout its superiority over the other solutions when it comes to data access performance from the cache. From what I can tell, bcache is unlike the previous solutions where it does not rely on the device-mapper framework and instead is a standalone module. At the time of this writing, it is set to be integrated into release 3.10 of the Linux kernel tree. Unfortunately, I haven't had the opportunity or the appropriate setup to test bcache. As a result, I haven't been able to dive any deeper into this solution and benchmark its performance.

EnhanceIO:

EnhanceIO is an SSD caching solution produced by STEC, Inc., and hosted on GitHub. It was greatly inspired by the work done by Facebook for FlashCache, and although it's open-source, a commercial version is offered by the company for those seeking additional support. STEC did not simply modify a few lines of code of FlashCache and republish it. Instead, STEC rewrote the write-back caching logic while also improving other areas, such as memory footprint, failure handling and more. As with bcache, I haven't had the opportunity to install and test EnhanceIO.

Summary

These solutions are intended to provide users with near SSD speeds and HDD capacities at a significantly reduced cost. From the data center to your home office, these solutions can be deployed almost anywhere. They also can be tuned to operate more appropriately in their intended environments. Some of them even offer a variety of caching algorithm options, such as Least Recently Used (LRU), Most Recently Used (MRU), hybrids of the two or just a simple first-in first-out (FIFO) caching scheme. The first three options can be expensive regarding performance, as they require the tracking of cached data sets for what has been accessed and how recently in order to determine whether to discard it. FIFO, however, functions as a circular buffer in which the oldest cached data set will be discarded first. With the exception of RapidCache, the SSD-focused modules also preserve metadata of the cache to ensure that any disruptions, including power cycles/outages, don't compromise the integrity of the data.

Petros Koutoupis is a full-time Linux kernel, device driver and application developer for embedded and server platforms. He has been working in the data storage industry for more than eight years and enjoys discussing the same technologies.

LJ Archive