Depth analysis of Apple Fusion Drive hybrid disk program

When the considerable effect moderate price SSD appears for the first time in 2008, the people feel that this technology is quite magical. With the passage of time, NAND and SSD improvement in price, but the price of mechanical hard or an order of magnitude lower than it. AnandTech's Anand Lal Shimpi said: I have been advocating a combination of SSD and HDD solutions. You can buy an SSD large enough to install the operating system, applications, and even one or two games and then loaded, and then everything else on HDD RAID-1 array. This method is used in the desktop, the effect is very good, but you have to manually manage files whereabouts. OEM is how to deal with this problem? I have always been curious. Because the the Education user's own hands, large, infrequently used files on a drive, and the other things to put on another drive, does not seem feasible. Intel 6 series chipset using Smart Response hybrid hard drive technology, a 20GB SLC SSD cache as a traditional hard drive, hoping to achieve a speed boost.

Since then, some other SSD caching solutions have also appeared. However, most of them used is inexpensive, small-capacity, of the mSATA SSD ineffective. Recently, OEM with SSD cache suppliers to barely meet the the certified minimum requirements of the ultra-extreme. Generally, this manner experience very poor. HDD manufacturers are trying to solve this problem, but they are a way to add a small amount of NAND mechanical hard drive. This usually makes the the HDD speed becomes faster, but with the SSD is still a far cry from the. NAND capacity storage devices has actually big enough to accommodate all of your data, the practice seems to be: you stick with 8MB L3 cache, in fact, you could have an extra few hundred dollars to buy a 16GB. Once you understand the benefits of the latter, the former seems to be meaningless compromise. Apple is one of the first to realize that this issue OEM. All only with NAND mainstream mobile devices (iPhone, iPad, and the MacBook Air). Recently, Apple's professional laptop even beginning to move in the direction of pure SSD forward (with Retina display MacBook Pro). Apple Mac does not participate in the capital of the low-price competition, to abandon HDD much easier. Even so, the 6 Mac (MBA, rMBP, MBP, Mac mini, iMac and Mac Pro), Apple is now shipping, only two default configuration without HDD. The rest are equipped with old-fashioned mechanical hard disk. IMac this device into a a pure SSD more difficult. Although ultra-extreme users (especially laptop users) have become accustomed to the equipment, do not have much storage space and a limited amount of storage used to desktops do not necessarily bear it.

Apple's approach this problem with other SSD / HDD hybrid solution is not much different. The difference is only in the SSD portion size as well as the software layer. Acquaintance Fusion Drive The new Mac mini and iMac can choose Apple's hybrid hard drive solutions Fusion Drive. It only has two versions: 1TB and 3TB. The 1TB only applies to iMac and upgrade the Mac mini ($ 799), 3TB applies only to the 27-inch iMac.

Fusion Drive is a 1TB or 3TB HDD (2.5-inch or 3.5-inch) with a 128GB Samsung PM830 SSD combination. In testing, I used a 27 inches iMac, 1TB of Fusion Drive. Apple Fusion Drive SSD cache capacity is much higher than the general program (most OEM 8GB to 24GB NAND). And Apple Fusion Drive selected Samsung PM830, SSD we tested in the best combination of performance and reliability. Although I personally prefer Link A Media or Intel S3700, after all, they have a good, stable performance, but for Apple, PM830 may be a more affordable option.

Of course, Fusion Drive previous hybrid / cache solutions, but the real difference is that the software part. Fusion Drive is not simply borrow Intel Smart Response Technology, it comes to OS X Logical Volume Manager Core Storage virtualization storage. The Core Storage earliest version of lions, which allows the operating system to treat multiple physical disks as one logical volume. Apple initially Core Storage to encrypt the entire disk, but now, the use of Core Storage has been expanded to the cougar Fusion Drive. Creating Fusion Drive is very simple. If you have multiple drives, you can use a simple terminal command to create a Fusion the Drive. If you buy one with a Fusion Drive Mac Apple automatically done for you all. Traditional SSD cache architecture, the Fusion not really cache the Drive, it will be based on the frequency of visits and the available space between SSD and HDD mobile data. A Fusion Drive capacity, in fact, is the sum of its components. In other words, a 1TB Fusion Drive capacity is actually 1TB + 128GB; 3TB Fusion Drive capacity is 3TB + 128GB.

The latest version of the "Disk Utility" (Disk Utility) Fusion Drive appears as a drive label Macintosh HD. But Apple did not try to hide its Fusion Drive Properties: system report, or use as third-party tools such as iStat Menus, you can see two drive situation:

128GB SSD capacity is 121.33GB. Note: Since OS X 10.6, Apple began radix display capacity of 10. Count you'll know how much space is used as a spare area:

Approximately 11.7% of the 128GiB NAND set aside as spare area, the standard Mac 128GiB SSD is the same, but this ratio is higher than the 6.7%. Increase the proportion of spare area help to improve the consistency of the performance, I personally hope that Samsung SSD can have 25% of the spare area. Boot Camp Fusion Drive, you can create additional partitions, these partitions are on the HDD. Fusion Drive Secret I tested with a new 27-inch iMac. In the beginning, I used a 128KB order of write operations (queue depth of 1), and use iStat Menus to monitor the status of the two drives, I found only SSD to receive the first write operation, there is no activity on the HDD. SSD write speed is 322MB / s. After writing to 117GB, HDD took over the write activity, the initial velocity of about 133 to 175MB / s.

Preliminary tests confirmed, Fusion Drive does use two drives. Written to the original 117GB SSD, 1TB next directly written to the HDD. I think that it is scheduled priority: Fusion Drive assume that SSD has enough free space (which will be discussed in detail), the first attempts to write to the SSD. Next, I would like to conduct random IO test, because this is the HDD SSD in terms of performance over the place, usually SSD cache or mixed programs fall short of the place. I first tried a worst case scenario: random write test covers all of the logical block address. In view of the total capacity of the Fusion Drive is 1.1TB, from this test, we can understand how Apple LBA (logical block address) mapping between the two drives. The results are interesting, but not too surprised. SSD and HDD write activities, but HDD some more (consume a greater proportion of the available LBA). Ordinary 4KB (QD16) random write test results about 0.51MB / s, it is subject to the restrictions of the Fusion Drive HDD. However, after the end of the mandate of the random write direct data movement between HDD and SSD. LBA is randomly selected, just the same (or similar space) address may be selected more than once, these logic blocks immediately be marked for upgrade to SSD. This is the first time I saw the Fusion Drive initiative to move data between drives. Full coverage for consumer-grade SSD random write test may be a bit unfair, is even more unfair for SSD / HDD hybrid solution. The Fusion Drive To learn the ability to deal with random IO, random write test the limit of the LBA 8GB. The results are very different. In the first round, the average speed of about 7 to 9MB / s, IO vast majority occur in the SSD, a small portion on the HDD. Three minutes after the test, I'm waiting for the Fusion Drive mobile data, and then repeated again. In the second round, the speed jumped 21.9MB / s, the more IO occurred in the SSD, the write activity on the HDD is still visible.

Figure: Most random write occurs on the SSD, but there is still part of the HDD to move a lot of data and after remapping LBA, almost all of the random write to the SSD, the speed is a lot faster In the third attempt, almost all of the random write on the SSD speeds of up to 98MB / s of peak debris is increased, the speed drops to a minimum 35MB / s. In my opinion, Apple seems to be based on the frequency of visits to LBA dynamically mapped to the SSD, which is a very positive performance boost. Ultimately, this is a significant difference between the usual SSD caching scheme and Fusion Drive. Most SSD caching solution seems to read the frequency-based Fusion Drive seems (at least in part) to consider which the LBA is frequently write and map them to the SSD.

It should be noted, that the subsequent random write test result is not the same. When I am filled with more data and applications Fusion Drive (to 80%) is filled with real data and applications, I did not see the random write performance to achieve such a high level again. In each run, I see a very short period of time the data is moved around, but the the random IO occurrence in HDD and SDD ratio is about 7:1. Given the difference in capacity between the two drives, this proportion has great significance. If your work load contains a large number of random writes, covering all of the available space, Fusion Drive may be less suitable for you. However, such workloads are mostly corporate users, so you should not be a problem. Data finesse explore Apple publicity Fusion Drive, when it comes to file and application-level data movement, but in real use, data can move 128KB file blocks between SSD and HDD. Fs_usage tools, you can see the inner workings of Apple Fusion Drive. File to 128KB of file blocks in the drive between the mobile, and this is determined by the frequency of use of the file blocks. Customer workloads is often the order of the (pseudo-random) in the worst case, we can say with more certainty that, if you are in a 128KB file blocks access an LBA you actually access the same space more more LBA. The data migration process seems to mainly occur during idle, although I have also seen in the IO load is lighter, the drive between migration activities.

The rapid migration trigger mechanism is very interesting. Once the file is copied or created, the application starts, or other IO activity is completed, immediately begin data migration between SSD and HDD. When you fill Fusion Drive, the amount of data moving between the SSD and the HDD shrink dramatically. The thing is supposed to be like this. Infrequently accessed data should be placed on the HDD, the really important things will stay on the SSD. When the Fusion Drive At the time of writing, Apple would not be so positive release of the data on the SSD.

The data migration process itself is very simple, the data is marked for the promotion and demotion, it is physically copied to new storage devices, and then it was removed. In the migration process, if there is a power failure, Fusion Drive will not cause any loss of data, the source block will be removed only when a 128KB file two copies of the block in place. Apple last year on it, but I only saw told me. The 128KB file blocks to move data between the HDD and SSD, Apple can write the SSD part defragmentation. The Fusion Drive write preferred destination SSD (less than 128KB file block, random and pseudo-random write), any migration from the HDD to the SSD but are large pieces written in the order, if the drive there is a lot of debris, which will trigger a garbage block recovery process. SSD performance will certainly be reduced over time, but the process helps to maintain its performance, because SSD is almost always been run in a filled condition, and that it will receive various Irrelevant write. I mentioned earlier, I hope Apple more PM830 NAND space reserved for the spare area. I suspect that Apple did not set aside more spare area for cost considerations. Application experience In real use, Fusion Drive can give us what kind of experience it? Put simply, it's effect is surprisingly good. SSD caching solution I used all the more like a HDD instead of SSD, but Apple Fusion Drive experience is almost in the middle of HDD and SSD. Almost all programs installed are first SSD, I feel really like SSD Fusion Drive. When installing the application, copy anything, is generally to go to SSD. Seems to 4GB after it began with the HDD. But with a little effort, you can make the Fusion Drive began after only 1 to 2GB written to the HDD. I created a sequence with Iometer test file, monitoring Fusion Drive when to stop the file is written to the SSD, the process stops, rename file, create a file and start again. The results are as shown below:

You can see, if your speed is fast enough, you may be less than 2GB let HDD take over the write operation. I do not know that this is not the SSD on the amount of space available, but likely is, it really does not make sense because 121GB of SSD without fully. Most of the time, when you do not actively write SSD, Fusion Drive can be maintained at least 4GB of SSD space available. Note: when you first use it, the Fusion Drive the majority are empty, you write almost anything, regardless of size, will directly into the SSD. However, when the capacity to withstand a certain pressure, Apple will change the method: any file up to write 4GB SSD, and the rest is put on the HDD. Install Apple's OS X Developer Tools Xcode itself, I can confirm that. The latter is close to 4GB, but the default situation, it's much to the SSD. The data generated by the application as well. I use Xcode to create Adium, a 682MB project, the entire compilation process occurs in the SSD - mechanical hard disk has been no activity. I'm trying to create a larger project nearly 2GB of Firefox. Then short HDD activity, but the vast majority of activities are on the SSD. I copied a lot of video files (> 10GB) to an iMac, now I have to copy it to a new position, and pay close attention to the movements of the Fusion Drive. The first 2GB is transferred from the SSD to SSD, the next 2GB is read from the HDD, write to the SSD. About 4GB be copied, the source drive and target drive into the HDD. In my opinion, the correct approach should be by default to all files into the HDD, unless they are accessed frequently. Apple's approach seems to be a reasonable compromise, but it is the way of file blocks into the SSD than I thought more positive. I repeated this test with another video files that I have never visited, and get completely different results. The entire file on the HDD. I repeated this test, I visited several times iPhoto library, to my surprise, most of the iPhoto library on the HDD, but also some little read from SSD. But in both cases, the copy of the target drive, of course, are the SSD. I Anandtech folder size exceeds 32GB, containing text, photos, presentations, benchmark results, and I have done almost all of the evaluation. Although this folder is very important, but the majority of it which are not frequently accessed. Copy this folder, I found almost entirely on the HDD. 38GB "My Documents" folder, too, most of this folder is not read. On the other hand, the application is almost always placed on the SSD. Before I often manually operating systems and applications on the SSD, a large media files on the HDD, Fusion Drive to automate this process, and the results were quite good. The only difference between the I manually collate information and automation Fusion Drive is my "My Documents" folder and Anandtech files on the SSD by default. Doing so not only between the performance considerations, but also to reliability, because the hard drive is easier than SSD go wrong. Fusion Drive Performance The Fusion Drive benchmark is a bit difficult, because it is always written to the SSD. If you do not fill Fusion Drive, you write large amounts of data may all be on the SSD. If you do fill it, but the test data set <4GB you measure SSD performance. I came up with a relatively simple use cases, can covers two drives. My Fusion Drive is now filled with 70%, which means that the capacity of SSD is about to reach the peak of (In addition to the the 4GB buffer). I exported as TIFF format photos iPhoto library contains 703 photos. The resulting files are pretty large, 297 photos, 4GB of SSD write buffer is full, all subsequent export photos to the HDD. I'll give this process time, and then contrast it with a HDD partition on the iMac, and also compared with a Samsung PM830 SSD connected via USB 3.0. Results somewhat biased in pure HDD configuration, because of the write operation is substantially the order of:

The figure accurately summed up my experience Fusion Drive: it is almost in the middle of HDD and SSD. In this test, the speed increase is not large, but this is mainly because we use a relatively low queue depth sequential transmission. For random workloads, Fusion Drive HDD gap will widen. Unfortunately, I can not find a good application use cases to generate the 4GB + pseudo random data, so that it can be repetitive high to the benchmark level. If I have been of Fusion Drive very large order of write operations (a file up to 260GB) Finally, it would be silly to swap, no longer migrate data to the SSD, unless you reboot the machine. I suspect that this bug is not triggered by a normal automatic test, but it does provide an interesting situation, and I can use it to carry out the test. Start of pre-installed applications I frequently use the iMac, they are still on the SSD, but some of the later installation of the program is not so. Especially Photoshop CS6, it is part of the SSD, some on the HDD. Fusion Drive for the work load is too large, the SSD can not be dealt with separately pseudo random read performance, this is a good benchmark. I tested Photoshop start-up time on the Fusion Drive with pure HDD partition, as well as via the USB 3.0 connection on PM830 compared. Compare results again reflect my experience:

Relative to the pure the HDD, Fusion Drive has significantly improved, 40% faster start-up time. However, pure SSD is also faster than half. Please note, if Photoshop is one of the most commonly used applications, it will be completely moved to the SSD. About the same performance with pure SSD. The test does not appear in this situation, because my 1.1TB Fusion Drive drive almost full 80%. The Fusion Drive the practical limitations Apple Fusion Drive write SSD is very positive, but your data is more, its algorithm seems to become more conservative. That was not very shocking, but it should be noted that lower total usage of the drive, I need to do almost everything using SSD, but when the application needs more than the Fusion Drive can easily cope with range, this platform for what can be moved to the SSD "become more picky. To remember, 128GB of storage space is enough to store all your most frequently used applications, data and operating systems, the Fusion Drive experience is more like a HDD. In order to simulate and prove it, I move my 200GB + MacBook Pro image to your iMac. Please note that most of this 200GB I frequently used applications and data. Test to the end, I think I need more SSD. Spotlight search takes longer than pure SSD, not all applications are instant start, add photos from Safari to iPhoto needs a long time little ...... Fusion Drive may be good, but it is not magic. If you really need more than 128GB over SSD, Fusion Drive may not be suitable for you. Conclusion Apple Fusion Drive is probably the best I've used SSD / HDD hybrid solution. This is not cutting-edge science, and it does a good SSD (Samsung PM830) a lot of NAND (128GB) and some very positive, intelligent software (Apple Core Storage LVM) combined. The Fusion Drive may not be considered innovative, but it's certainly correct. If you do not have a large number of applications and data, media files is not too much, then for you, Fusion Drive drive should be a reasonable SSD plus HDD configuration, the media files on the HDD, all things are frequently used in SSD on. Although I very much want a larger cache, but I believe that, for most workloads and usage patterns, 128GB enough good. May be required for me personally, Fusion Drive 256GB cache in order to win my heart, but I also understand that I do not necessarily its target audience.

The real question is: whether it is good value for money. Personally, I certainly hope that all SSD, media files on HDD array manually. The Fusion Drive did not do anything to reduce the possibility of "HDD may be earlier than the SSD go wrong", if you use the built-in SSD (Thunderbolt or USB 3.0) HDD RAID array connected, you will master the greater flexibility sex. Unfortunately, Apple does not always provide pure SSD version of the iMac. The Apple iMac limit your choice. For the 21.5-inch model, you can either buy pure HDD version, Fusion Drive version can also be purchased, but there is no pure SSD version. I think you should certainly choose the Fusion Drive version. If you do not intend to buy their own SSD to replace its built-in HDD, Fusion Drive is definitely a right choice. For the 27-inch iMac, Apple provides a pure SSD version, but its capacity is only 768GB, the upgrade price of 1300 U.S. dollars. Suddenly, the $ 250 Fusion Drive upgrade sounds more reasonable. Boldface in the following figure, I recommend the optional program. Please note that if you do not intend to own DIY (if you decide to DIY, then buy the biggest SSD you can find to put those used with external large-capacity memory):

When NAND becomes cheap enough and large enough, iMac completely abandoned HDD it? Or computer will eventually be enough NAND, everything outside in addition to large media files can be stored in it? Year, Apple could double the NAND Fusion Drive, while the cost remains the same. If you see farther in the post-NAND era, memory hierarchy (memory hierarchy) may appear some truly revolutionary things.