Continuing to examine input/output acceleration technologies as applied to storage systems, initiated in , we cannot overlook a very popular option known as auto tiering. Although the working ideology of this feature is quite similar across various storage system manufacturers, we will explore the specifics of auto tiering implementation using the example of .
Despite the diversity of data stored on storage systems, this data can be divided into several groups based on their demand (frequency of use). For the most popular ('hot') data, it is critically important to ensure maximum access speed, while the processing of less demanded ('cold') data can be done with lower priority.
The functionality of auto tiering is precisely for organizing such a scheme. In this case, the data array consists not of homogeneous disks, but of several groups of drives forming different storage tiers. A special algorithm automatically moves data between levels to ensure maximum overall performance.
NAS support up to three storage levels:
- Tier 1: SSD, maximum performance
- Tier 2: HDD SAS 10K/15K, high performance
- Tier 3: HDD NL-SAS 7.2K, maximum capacity
An auto tiering pool can contain all three levels or only two in any combinations. Within each tier, drives are combined into familiar RAID groups. For maximum flexibility, the RAID level in each tier can be different. That is, for example, nothing prevents organizing a structure such as 4x SSD RAID10 + 6x HDD 10K RAID5 + 12 HDD 7.2K RAID6.
After creating volumes (virtual disks) on pool, background statistics collection begins on all input/output operations. For this purpose, space is 'sliced' into 1GB blocks (so-called sub LUNs). Each time a block is accessed, it is assigned a coefficient of 1. Then, over time, this coefficient decreases. After 24 hours, with no input/output requests to this block, it will already be equal to 0.5 and will continue to fall with each subsequent hour.
At a certain point in time (by default, every day at midnight), the collected results are ranked based on sub LUN activity according to their coefficients. Based on this, a decision is made about which blocks to move and in which direction. After that, the actual data relocation between levels takes place.
In the Qsan storage system, the management of the tiering process is excellently implemented using various parameters, allowing for highly flexible tuning of the final performance of the array.
To determine the initial data location and the preferred direction for their movement, policies are set separately for each volume:
- the auto tiering – default policy, where the initial placement and movement direction are defined automatically, meaning that 'hot' data tend to move to the highest level while 'cold' data shift downwards. The initial placement is chosen based on available space at each level. However, it is essential to understand that the system primarily aims to maximize the use of the fastest storage devices. Therefore, when free space is available, data will be placed on the upper levels. This policy is suitable for most scenarios where data demand cannot be predicted in advance.
- Start high, then Auto Tiering – the only difference from the previous one is the initial data placement (at the fastest level)
- Maximum high level – data always strives to occupy the fastest level. If they are moved down during operation, they will be relocated back at the first opportunity. This policy is ideal for data requiring the fastest access.
- Minimum level – data always strives to occupy the lowest level. This policy is perfect for rarely used data (for example, archives).
- No movement – the system automatically determines the initial location of the data and does not move it. However, statistics continue to be collected in case relocation is required later.
It is worth noting that although policies are defined at the creation of each volume, they can be changed 'on the fly' multiple times during the system's lifecycle.
In addition to policies for the tiering mechanism, the frequency and speed of data movement between tiers can also be configured. You can specify a specific move time: daily or on certain days of the week, as well as reduce the statistics collection interval to a few hours (the minimum frequency is 2 hours). If there is a need to limit the time for data moving operations, you can set a time frame (window for relocation). Additionally, the speed of relocation is specified – 3 modes: fast, medium, slow.
In case of an urgent need for data relocation, it can be performed manually at any time upon the administrator's command.
It is clear that the more frequently and quickly data is moved between tiers, the more flexible the storage system will adjust to current operational conditions. However, it is also important to remember that moving data adds additional load (primarily on the disks), so it is not advisable to 'race' data without extreme necessity. It's better to schedule moves during times of minimal load. If the storage system constantly requires high performance in a 24/7 mode, then it is advisable to reduce the pace of relocation to a minimum.
The abundance of tiering settings will undoubtedly please advanced users. However, there is nothing to worry about for those encountering this technology for the first time. You can fully trust the default settings (Auto Tiering policy, moving at maximum speed once a night) and adjust certain parameters as you accumulate statistics to achieve the desired result.
Comparing tiering with another equally popular performance enhancement technology, such as , it is important to remember the different principles of how their algorithms work.
SSD caching
the auto tiering
Speed of achieving effect
Almost instantly. But the noticeable effect only after 'warming up' the cache (minutes to hours)
After collecting statistics (from 2 hours, ideally – 24 hours) plus time for data relocation
Duration of the effect
Until the data is replaced by a new batch (minutes to hours)
While the demand for the data remains relevant (24 hours or more)
Indications for use
Instant performance boost for a short duration (databases, virtualization environments)
Sustained performance increase over a long period (file, web, mail servers)
One of the features of tiering is the ability to use it not only for scenarios like 'SSD + HDD', but also 'fast HDD + slow HDD' or even all three levels, which would be impossible with SSD caching.
Testing
To test the operation of tiering algorithms, we conducted a simple test. A pool was created from two levels of SSD (RAID 1) + 7.2K HDD (RAID 1), on which a volume was placed with a 'minimum level' policy. That is, data must always be located on the slower disks.
The management interface clearly shows the distribution of data between the levels.
After filling the volume with data, we switched the placement policy to Auto Tiering and started an IOmeter test.
After several hours of testing, when the system was able to accumulate statistics, the relocation process began.
At the end of the data movement, our test volume fully 'migrated' to the upper level (SSD).
Verdict
Auto Tiering is a remarkable technology that allows you to enhance storage system performance with minimal material and time costs by utilizing high-speed drives more intensively. Applicable to the only investment is a one-time license purchase without limitations on the volume/number of disks/racks/etc. This functionality is equipped with such rich settings that it can satisfy almost any business task. The visualization of processes in the interface will allow effective device management.
Source: habr.com
