In our blog on Habré we not only tell about development of a cloud service 1cloud, but also we write about new technologies much, including in the field of processors and memory. Today we present to yours adapted translation of a note of the director of research of the IDC company Phil Goodwin (Phil Goodwin) about pluses and minuses of technologies of hybrid arrays of data storage and arrays a flash memory.
With falling of the prices one SSD more and more companies offer the arrays entirely constructed on a flash memory, but whether really they it is better than the hybrid arrays containing both solid-state drives, and hard drives?
For comparison of different technology solutions we often address a ratio the price/performance, however emergence of SSD led to the fact what buyers had to select that it is more important for them. A flash memory offered magnificent speeds and very high cost of gigabyte. On the other hand, multiterabyte hard drives (HDD) not so strongly devastated purses, but, having speed in 75 IOPS, could not cope with a heavy load.
Gigabyte cost on the hard drive is less, than at SSD, but it loses in work speed.
Hybrid arrays had to become a compromise. Adding a flash memory in the amount of 2-5% of the total amount of memory, it is possible to double quantity of IOPS practically that will reduce latency of system to the level of 3-5 ms (without flash technologies it makes more than 10 ms). Thus, having paid only 10-20% more, you have an opportunity to double array performance.
As if digits in 3-5 ms looked good, in some situations of it it appears insufficiently. All-flash-arrays are capable to provide latency less than a millisecond at due QoS – quality of service. In such architecture there are no delays because of data sampling as, technically, input-output operations are executed from a cache memory. Only think, decrease in latency from 10 ms to 5 ms, and then to 0,5 ms is a twentyfold increase in speed of work in comparison with hard drives. Today the prices of SSD and HDD very close approached to each other, but, despite it, the majority of the organizations still are not able to afford to purchase hundreds terabyte a flash memory. If nevertheless the flash technology is required to implement the companies, then the knowledge of the main features of both approaches will help heads to make the weighed decision.
Five reasons to select a hybrid array or an all-flash-array
|It is better to use a hybrid array when||the all-flash-array is necessary when|
operational load are unknown
|It is required
latency less than a millisecond
|It is necessary
to optimize cost
|It is required
the guaranteed quality of data exchange
in indicators of latency is not an obstacle
|It is necessary
to optimize performance
transition from classical arrays
more important than cost
does not allow to perform complete transition to SSD, but it is required to increase
an opportunity to use little-known high-performance HDD that
to reduce the aggregate value of ownership of system
Ratio price/performance hybrid and all-flash-arrays
As the price and performance are two most important metrics upon purchase of writer devices, let's consider them at first. It is considered that the cost of purchase of a HDD array makes about 20% of the cost of its operation within three years. All-flash-arrays destroy this model: the cost of purchase grows, however operational costs are cut. For example, the Nimbus Data company, the vendor of all-flash-arrays, declares that their devices consume only 8 watts on each terabyte while HDD can consume up to 80 watts. Moreover, Nimbus says that users can fill a rack for 90%, and it will not cause decline in production – thus, the free space is released, and energy costs are optimized. The high-capacious disks SATA offer space gigabytes for information storage, but possess mediocre indicators of IOPS and latency. Flash memory, on the contrary, it is very similar to a cache and has the small volume, but high performance. For this reason it is worth comparing a total cost of ownership of hybrid arrays and all - flash - systems.
For estimation of cost the Hewlett-Packard company offers the following metrics: dollar/operation. This metrics reflects real operational costs while the dollar/GB and dollar/IOPS allow to estimate only cost of purchase. Still the ratio dollar/operation does not depend on carrier type. If it is required to execute a large number of operations, then the metrics dollar/operation will be less when using all-flash-arrays. For hybrid arrays a ratio dollar/operation will be less at a small amount of operations. Anyway, the result is predictable and is easily calculated.
Memory devices (their purchase can be considered as variable costs) can be considered by IT managers as the factors influencing the user experience. If users consider that the application works insufficiently quickly, they, most likely, will have a negative impression about all organization or provider. If the passenger of the airplane sees coffee spots on the folding little table, then he will begin to doubt also as service of engines. Perhaps, it makes sense to spend slightly more money to change opinion of the client to the best, especially during an era of IT outsourcing?
The deduplication and compression are used practically in all storage systems and allow to reduce the cost of one gigabyte of memory. However here it is necessary to make the remark. The deduplication and SIS – a component of storage of the only copies – are used at the level of OS and the level of applications (for example, the solutions VMware and SIS for Exchange). The deduplication and compression are not applied twice to one data set therefore administrators of storage systems can observe depreciation of gigabyte one value smaller than estimated.
Until recently for commercial computing applications performance of storages was evaluated only on IOPS. Perhaps, it was connected with the fact that arrays in principle could not guarantee certain indicators of QoS. All-flash-arrays changed rules of the game, than the SolidFire company used. The SolidFire technology allows to predict loading (IOPS) of separate sectors of an array that allows to select resources for reading/data record when it is necessary. Still it solves a problem of "the noisy neighbor". For example, in a typical hybrid array of the application compete with each other for section a flash memory (access is got by those who has more rights) that conducts to an overload and non-optimal performance. Fine tuning of IOPS for each application separately solves this problem.
So to select?
All famous vendors of storage systems already added the HDD solutions a flash technologies: hybrid devices or all-flash-elements. IBM, EMC Corp., NetApp Inc. and HP propose all above-mentioned solutions, however features of their execution seriously differ.
IBM expanded possibilities of the technology of virtualization of data storages of SAN Volume Controller (SVC) to manage all - flash - FlashSystem devices or hybrid arrays of lines of XIV Storage System, Strowize and DS. As all these devices can be controlled with the help of SVC, IBM applies the hybrid concept not only to arrays, but also to all ecosystem of storage of corporation. They want to receive the maximum flexibility at implementation of both types of carriers and combinations of arrays that will give more opportunities for scaling and accumulation of performance.
VMAX from EMC can virtualize both hybrid, and all-flash-solutions in one ecosystem. VMAX, VNX, VNXe and Isilon can also be both hybrid, and all-flash-products. Thus, users can decide that now it is more necessary to them. The company provides one more all-flash-solution – XtremIO. As there is no EMC need to separate hybrid and the all-flash-sentence, she can deliver traditional HDD arrays to buyers with the limited budget, hybrid configurations – to those who need performance, and all-flash-solutions if the client needs the guaranteed quality of service (QoS). The choice of a line depends on three indicators: reliability, availability and convenience of service. For example, the VMAX system is capable to provide availability at the "six nine" level (99,9999%) regardless of type.
HP offers 3PAR data storage systems constructed as on HDD, and a flash. The last possess interesting features: in one array flash cards with single-level and multi-level cells are located. The company recommends to use the first for data record, and the second for reading. All this is constructed on the basis of the Adaptive Optimization sub-LUN tiering technology. The built-in manager of cluster volumes has OS 3PAR and is able to virtualize all types of carriers, including new.
The NetApp company offers a flash for all arrays of FAS, V-series and E-series families. EF540 is all - flash - system. NetApp usually recommends to use Flash Pools of 1-2% of the total amount of memory. According to the company, for work with e-mails, web service, applications programming and cooperations will approach hybrid configurations. In these cases less than 150 000 IOPS are required, and latency in 3-5 ms is not an obstacle. At loading with a large number of operations or where latency less than a millisecond is required, it is better to use all-flash-arrays.
ZFS Storage ZS3 from Oracle are hybrid arrays, but the company says that users can reach 90% of probability of "hit" in a cache and latencies less than a millisecond by means of their flash/DRAM-architecture. Moreover, solutions of the company have some advantage in cost, in comparison with competitors. Users receive all performance - flash - systems at the price hybrid. Hybrid Storage Pool from Oracle automatically moves data on DRAM, reads from a flash memory and writes to a flash memory to optimize array performance.
Use of RAID
Data on a flash devices can also be protected by means of RAID, as in any other storage. Though most of suppliers support standard RAID techniques for a flash devices, use of RAID can cause value increase of information storage and reduction in the rate of its processing the same as in a case with HDD. IBM solves this problem with the help of Variable Stripe RAID in the FlashSystem technology where there is a built-in RAID 5 controller.
As a result RAID with a disk of parity and high data transmission rate where loading is distributed between several controllers turned out. NetApp began to use Dynamic Disk Pools ("dynamic disk pools") in the SANtricity (E-series). Thanks to this technology data are distributed evenly on all disks. NetApp claims that the system allows to recover quickly faulty disks, having saved at the same time high performance. Nimbus uses the algorithm of RAID 5 developed for the purpose of decrease in degree of wear of the equipment which is shown at record in a flash memory.
Despite distinctions in approaches and architecture, the companies suppliers agree in two things. First, if you need latency less than a millisecond or QoS, then use all-flash-arrays. In a case with Oracle it is possible to use the hybrid array capable to provide practically the same performance. The applications demanding QoS include [for example] solutions from area of electronic commerce where quality of the user experience, and analytical products in the field of decision-making support where one of the major parameters is work speed is regarded as of paramount importance.
Secondly, if loading is unknown or changes over time, then it is necessary to use hybrid devices which will cope with objectives, but for smaller cost. Nimbus, SolidFire, and also other all-flash-suppliers propose own solutions for this case – their products can have interesting characteristics the price/performance and the low cost of IOPS. The main scopes are e-mail and applications for [organization] of collaboration – everything where instant access to a large number of data is not required.
This article is a translation of the original post at habrahabr.ru/post/272899/
If you have any questions regarding the material covered in the article above, please, contact the original author of the post.
If you have any complaints about this article or you want this article to be deleted, please, drop an email here: firstname.lastname@example.org.
We believe that the knowledge, which is available at the most popular Russian IT blog habrahabr.ru, should be accessed by everyone, even though it is poorly translated.
Shared knowledge makes the world better.