1.\" 2.\" swapcache - Cache clean filesystem data & meta-data on SSD-based swap 3.\" 4.\" Redistribution and use in source and binary forms, with or without 5.\" modification, are permitted provided that the following conditions 6.\" are met: 7.\" 1. Redistributions of source code must retain the above copyright 8.\" notice, this list of conditions and the following disclaimer. 9.\" 2. Redistributions in binary form must reproduce the above copyright 10.\" notice, this list of conditions and the following disclaimer in the 11.\" documentation and/or other materials provided with the distribution. 12.Dd February 7, 2010 13.Dt SWAPCACHE 8 14.Os 15.Sh NAME 16.Nm swapcache 17.Nd a mechanism to use fast swap to cache filesystem data and meta-data 18.Sh SYNOPSIS 19.Cd sysctl vm.swapcache.accrate=100000 20.Cd sysctl vm.swapcache.maxfilesize=0 21.Cd sysctl vm.swapcache.maxburst=2000000000 22.Cd sysctl vm.swapcache.curburst=4000000000 23.Cd sysctl vm.swapcache.minburst=10000000 24.Cd sysctl vm.swapcache.read_enable=0 25.Cd sysctl vm.swapcache.meta_enable=0 26.Cd sysctl vm.swapcache.data_enable=0 27.Cd sysctl vm.swapcache.use_chflags=1 28.Cd sysctl vm.swapcache.maxlaunder=256 29.Cd sysctl vm.swapcache.hysteresis=(vm.stats.vm.v_inactive_target/2) 30.Sh DESCRIPTION 31.Nm 32is a system capability which allows a solid state disk (SSD) in a swap 33space configuration to be used to cache clean filesystem data and meta-data 34in addition to its normal function of backing anonymous memory. 35.Pp 36Sysctls are used to manage operational parameters and can be adjusted at 37any time. 38Typically a large initial burst is desired after system boot, 39controlled by the initial 40.Va vm.swapcache.curburst 41parameter. 42This parameter is reduced as data is written to swap by the swapcache 43and increased at a rate specified by 44.Va vm.swapcache.accrate . 45Once this parameter reaches zero write activity ceases until it has 46recovered sufficiently for write activity to resume. 47.Pp 48.Va vm.swapcache.meta_enable 49enables the writing of filesystem meta-data to the swapcache. 50Filesystem 51metadata is any data which the filesystem accesses via the disk device 52using buffercache. 53Meta-data is cached globally regardless of file or directory flags. 54.Pp 55.Va vm.swapcache.data_enable 56enables the writing of clean filesystem file-data to the swapcache. 57Filesystem filedata is any data which the filesystem accesses via a 58regular file. 59In technical terms, when the buffer cache is used to access 60a regular file through its vnode. 61Please do not blindly turn on this option, see the 62.Sx PERFORMANCE TUNING 63section for more information. 64.Pp 65.Va vm.swapcache.use_chflags 66enables the use of the 67.Va cache 68and 69.Va noscache 70.Xr chflags 1 71flags to control which files will be data-cached. 72If this sysctl is disabled and 73.Va data_enable 74is enabled, the system will ignore file flags and attempt to 75swapcache all regular files. 76.Pp 77.Va vm.swapcache.read_enable 78enables reading from the swapcache and should be set to 1 for normal 79operation. 80.Pp 81.Va vm.swapcache.maxfilesize 82controls which files are to be cached based on their size. 83If set to non-zero only files smaller than the specified size 84will be cached. 85Larger files will not be cached. 86.Pp 87.Va vm.swapcache.maxlaunder 88controls the maximum number of clean VM pages which will be added to 89the swap cache and written out to swap on each poll. 90Swapcache polls ten times a second. 91.Pp 92.Va vm.swapcache.hysteresis 93controls how many pages swapcache waits to be added to the inactive page 94queue before continuing its scan. 95Once it decides to scan it continues subject to the above limitations 96until it reaches the end of the inactive page queue. 97This parameter is designed to make swapcache generate more bulky bursts 98to swap which helps SSDs reduce write amplification effects. 99.Sh PERFORMANCE TUNING 100Best operation is achieved when the active data set fits within the 101swapcache. 102.Pp 103.Bl -tag -width 4n -compact 104.It Va vm.swapcache.accrate 105This specifies the burst accumulation rate in bytes per second and 106ultimately controls the write bandwidth to swap averaged over a long 107period of time. 108This parameter must be carefully chosen to manage the write endurance of 109the SSD in order to avoid wearing it out too quickly. 110Even though SSDs have limited write endurance, there is massive 111cost/performance benefit to using one in a swapcache configuration. 112.Pp 113Let's use the old Intel X25V 40GB MLC SATA SSD as an example. 114This device has approximately a 11540TB (40 terabyte) write endurance, but see later 116notes on this, it is more a minimum value. 117Limiting the long term average bandwidth to 100KB/sec leads to no more 118than ~9GB/day writing which calculates approximately to a 12 year endurance. 119Endurance scales linearly with size. 120The 80GB version of this SSD 121will have a write endurance of approximately 80TB. 122.Pp 123MLC SSDs have a 1000-10000x write endurance, while the lower density 124higher-cost SLC SSDs have a 10000-100000x write endurance, approximately. 125MLC SSDs can be used for the swapcache (and swap) as long as the system 126manager is cognizant of its limitations. 127However, over the years tests have shown the SLC SSDs do not really live 128up to their hype and are no more reliable than MLC SSDs. Instead of 129worrying about SLC vs MLC, just use MLC (or TLC or whateve), leave 130more space unpartitioned which the SSD can utilize to improve durability, 131and be cognizant of the SSDs rate of wear. 132.Pp 133.It Va vm.swapcache.meta_enable 134Turning on just 135.Va meta_enable 136causes only filesystem meta-data to be cached and will result 137in very fast directory operations even over millions of inodes 138and even in the face of other invasive operations being run 139by other processes. 140.Pp 141For 142.Nm HAMMER 143filesystems meta-data includes the B-Tree, directory entries, 144and data related to tiny files. 145Approximately 6 GB of swapcache is needed 146for every 14 million or so inodes cached, effectively giving one the 147ability to cache all the meta-data in a multi-terabyte filesystem using 148a fairly small SSD. 149.Pp 150.It Va vm.swapcache.data_enable 151Turning on 152.Va data_enable 153(with or without other features) allows bulk file data to be cached. 154This feature is very useful for web server operation when the 155operational data set fits in swap. 156However, care must be taken to avoid thrashing the swapcache. 157In almost all cases you will want to leave chflags mode enabled 158and use 'chflags cache' on governing directories to control which 159directory subtrees file data should be cached for. 160.Pp 161DragonFly uses generously large kern.maxvnodes values, 162typically in excess of 400K vnodes, but large numbers 163of small files can still cause problems for swapcache. 164When operating on a filesystem containing a large number of 165small files, vnode recycling by the kernel will cause related 166swapcache data to be lost and also cause the swapcache to 167potentially thrash. 168Cache thrashing due to vnode recyclement can occur whether chflags 169mode is used or not. 170.Pp 171To solve the thrashing problem you can turn on HAMMER's 172double buffering feature via 173.Va vfs.hammer.double_buffer . 174This causes HAMMER to cache file data via its block device. 175HAMMER cannot avoid also caching file data via individual vnodes 176but will try to expire the second copy more quickly (hence 177why it is called double buffer mode), but the key point here is 178that 179.Nm 180will only cache the data blocks via the block device when 181double_buffer mode is used and since the block device is associated 182with the mount, vnode recycling will not mess with it. 183This allows the data for any number (potentially millions) of files to 184be swapcached. 185You still should use chflags mode to control the size of the dataset 186being cached to remain under 75% of configured swap space. 187.Pp 188Data caching is definitely more wasteful of the SSD's write durability 189than meta-data caching. 190If not carefully managed the swapcache may exhaust its burst and smack 191against the long term average bandwidth limit, causing the SSD to wear 192out at the maximum rate you programmed. 193Data caching is far less wasteful and more efficient 194if you provide a sufficiently large SSD. 195.Pp 196When caching large data sets you may want to use a medium-sized SSD 197with good write performance instead of a small SSD to accommodate 198the higher burst write rate data caching incurs and to reduce 199interference between reading and writing. 200Write durability also tends to scale with larger SSDs, but keep in mind 201that newer flash technologies use smaller feature sizes on-chip 202which reduce the write durability of the chips, so pay careful attention 203to the type of flash employed by the SSD when making durability 204assumptions. 205For example, an Intel X25-V only has 40MB/s in write performance 206and burst writing by swapcache will seriously interfere with 207concurrent read operation on the SSD. 208The 80GB X25-M on the otherhand has double the write performance. 209Higher-capacity and larger form-factor SSDs tend to have better 210write-performance. 211But the Intel 310 series SSDs use flash chips with a smaller feature 212size so an 80G 310 series SSD will wind up with a durability relative 213close to the older 40G X25-V. 214.Pp 215When data caching is turned on you can fine-tune what gets swapcached 216by also turning on swapcache's chflags mode and using 217.Xr chflags 1 218with the 219.Va cache 220flag to enable data caching on a directory-tree (recursive) basis. 221This flag is tracked by the namecache and does not need to be 222recursively set in the directory tree. 223Simply setting the flag in a top level directory or mount point 224is usually sufficient. 225However, the flag does not track across mount points. 226A typical setup is something like this: 227.Pp 228.Dl chflags cache /etc /sbin /bin /usr /home 229.Dl chflags noscache /usr/obj 230.Pp 231It is possible to tell 232.Nm 233to ignore the cache flag by leaving 234.Va vm.swapcache.use_chflags 235set to zero. 236In many situations it is convenient to simply not use chflags mode, but 237if you have numerous mixed SSDs and HDDs you may want to use this flag 238to enable swapcache on the HDDs and disable it on the SSDs even if 239you do not care about fine-grained control. 240.Nm chflag Ns 'ing . 241.Pp 242Filesystems such as NFS which do not support flags generally 243have a 244.Va cache 245mount option which enables swapcache operation on the mount. 246.Pp 247.It Va vm.swapcache.maxfilesize 248This may be used to reduce cache thrashing when a focus on a small 249potentially fragmented filespace is desired, leaving the 250larger (more linearly accessed) files alone. 251.Pp 252.It Va vm.swapcache.minburst 253This controls hysteresis and prevents nickel-and-dime write bursting. 254Once 255.Va curburst 256drops to zero, writing to the swapcache ceases until it has recovered past 257.Va minburst . 258The idea here is to avoid creating a heavily fragmented swapcache where 259reading data from a file must alternate between the cache and the primary 260filesystem. 261Doing so does not save disk seeks on the primary filesystem 262so we want to avoid doing small bursts. 263This parameter allows us to do larger bursts. 264The larger bursts also tend to improve SSD performance as the SSD itself 265can do a better job write-combining and erasing blocks. 266.Pp 267.It Va vm_swapcache.maxswappct 268This controls the maximum amount of swapspace 269.Nm 270may use, in percentage terms. 271The default is 75%, leaving the remaining 25% of swap available for normal 272paging operations. 273.El 274.Pp 275It is important to ensure that your swap partition is nicely aligned. 276The standard DragonFly 277.Xr disklabel 8 278program guarantees high alignment (~1MB) automatically. 279Swap-on HDDs benefit because HDDs tend to use a larger physical sector size 280than 512 bytes, and proper alignment for SSDs will reduce write amplification 281and write-combining inefficiencies. 282.Pp 283Finally, interleaved swap (multiple SSDs) may be used to increase 284swap and swapcache performance even further. 285A single SATA-II SSD is typically capable of reading 120-220MB/sec. 286Configuring two SSDs for your swap will 287improve aggregate swapcache read performance by 1.5x to 1.8x. 288In tests with two Intel 40GB SSDs 300MB/sec was easily achieved. 289With two SATA-III SSDs it is possible to achieve 600MB/sec or better 290and well over 400MB/sec random-read performance (versus the ~3MB/sec 291random read performance a hard drive gives you). 292Faster SATA interfaces or newer NVMe technologies have significantly 293more read bandwidth (3GB/sec+ for NVMe), but may still lag on the 294write bandwidth. 295With newer technologies, one swap device is usually plenty. 296.Pp 297.Dx 298defaults to a maximum of 512G of configured swap. 299Keep in mind that each 1GB of actually configured swap requires 300approximately 1MB of wired ram to manage. 301.Pp 302In addition there will be periods of time where the system is in 303steady state and not writing to the swapcache. 304During these periods 305.Va curburst 306will inch back up but will not exceed 307.Va maxburst . 308Thus the 309.Va maxburst 310value controls how large a repeated burst can be. 311Remember that 312.Va curburst 313dynamically tracks burst and will go up and down depending. 314.Pp 315A second bursting parameter called 316.Va vm.swapcache.minburst 317controls bursting when the maximum write bandwidth has been reached. 318When 319.Va minburst 320reaches zero write activity ceases and 321.Va curburst 322is allowed to recover up to 323.Va minburst 324before write activity resumes. 325The recommended range for the 326.Va minburst 327parameter is 1MB to 50MB. 328This parameter has a relationship to 329how fragmented the swapcache gets when not in a steady state. 330Large bursts reduce fragmentation and reduce incidences of 331excessive seeking on the hard drive. 332If set too low the 333swapcache will become fragmented within a single regular file 334and the constant back-and-forth between the swapcache and the 335hard drive will result in excessive seeking on the hard drive. 336.Sh SWAPCACHE SIZE & MANAGEMENT 337The swapcache feature will use up to 75% of configured swap space 338by default. 339The remaining 25% is reserved for normal paging operations. 340The system operator should configure at least 4 times the SWAP space 341versus main memory and no less than 8GB of swap space. 342A typical 128GB SSD might use 64GB for boot + base and 56GB for 343swap, with 8GB left unpartitioned. The system might then have a large 344additional hard drive for bulk data. 345Even with many packages installed, 64GB is comfortable for 346boot + base. 347.Pp 348When configuring a SSD that will be used for swap or swapcache 349it is a good idea to leave around 10% unpartitioned to improve 350the SSDs durability. 351.Pp 352You do not need to use swapcache if you have no hard drives in the 353system, though in fact swapcache can help if you use NFS heavily 354as a client. 355.Pp 356The 357.Va vm_swapcache.maxswappct 358sysctl may be used to change the default. 359You may have to change this default if you also use 360.Xr tmpfs 5 , 361.Xr vn 4 , 362or if you have not allocated enough swap for reasonable normal paging 363activity to occur (in which case you probably shouldn't be using 364.Nm 365anyway). 366.Pp 367If swapcache reaches the 75% limit it will begin tearing down swap 368in linear bursts by iterating through available VM objects, until 369swap space use drops to 70%. 370The tear-down is limited by the rate at 371which new data is written and this rate in turn is often limited by 372.Va vm.swapcache.accrate , 373resulting in an orderly replacement of cached data and meta-data. 374The limit is typically only reached when doing full data+meta-data 375caching with no file size limitations and serving primarily large 376files, or bumping 377.Va kern.maxvnodes 378up to very high values. 379.Sh NORMAL SWAP PAGING ACTIVITY WITH SSD SWAP 380This is not a function of 381.Nm 382per se but instead a normal function of the system. 383Most systems have 384sufficient memory that they do not need to page memory to swap. 385These types of systems are the ones best suited for MLC SSD 386configured swap running with a 387.Nm 388configuration. 389Systems which modestly page to swap, in the range of a few hundred 390megabytes a day worth of writing, are also well suited for MLC SSD 391configured swap. 392Desktops usually fall into this category even if they 393page out a bit more because swap activity is governed by the actions of 394a single person. 395.Pp 396Systems which page anonymous memory heavily when 397.Nm 398would otherwise be turned off are not usually well suited for MLC SSD 399configured swap. 400Heavy paging activity is not governed by 401.Nm 402bandwidth control parameters and can lead to excessive uncontrolled 403writing to the SSD, causing premature wearout. 404This isn't to say that 405.Nm 406would be ineffective, just that the aggregate write bandwidth required 407to support the system might be too large to be cost-effective for a SSD. 408.Pp 409With this caveat in mind, SSD based paging on systems with insufficient 410RAM can be extremely effective in extending the useful life of the system. 411For example, a system with a measly 192MB of RAM and SSD swap can run 412a -j 8 parallel build world in a little less than twice the time it 413would take if the system had 2GB of RAM, whereas it would take 5x to 10x 414as long with normal HDD based swap. 415.Sh USING SWAPCACHE WITH NORMAL HARD DRIVES 416Although 417.Nm 418is designed to work with SSD-based storage it can also be used with 419HD-based storage as an aid for offloading the primary storage system. 420Here we need to make a distinction between using RAID for fanning out 421storage versus using RAID for redundancy. There are numerous situations 422where RAID-based redundancy does not make sense. 423.Pp 424A good example would be in an environment where the servers themselves 425are redundant and can suffer a total failure without effecting 426ongoing operations. When the primary storage requirements easily fit onto 427a single large-capacity drive it doesn't make a whole lot of sense to 428use RAID if your only desire is to improve performance. If you had a farm 429of, say, 20 servers supporting the same facility adding RAID to each one 430would not accomplish anything other than to bloat your deployment and 431maintenance costs. 432.Pp 433In these sorts of situations it may be desirable and convenient to have 434the primary filesystem for each machine on a single large drive and then 435use the 436.Nm 437facility to offload the drive and make the machine more effective without 438actually distributing the filesystem itself across multiple drives. 439For the purposes of offloading while a SSD would be the most effective 440from a performance standpoint, a second medium sized HD with its much lower 441cost and higher capacity might actually be more cost effective. 442.Sh EXPLANATION OF STATIC VS DYNAMIC WEARING LEVELING, AND WRITE-COMBINING 443Modern SSDs keep track of space that has never been written to. 444This would also include space freed up via TRIM, but simply not 445touching a bit of storage in a factory fresh SSD works just as well. 446Once you touch (write to) the storage all bets are off, even if 447you reformat/repartition later. It takes sending the SSD a 448whole-device TRIM command or special format command to take it back 449to its factory-fresh condition (sans wear already present). 450.Pp 451SSDs have wear leveling algorithms which are responsible for trying 452to even out the erase/write cycles across all flash cells in the 453storage. The better a job the SSD can do the longer the SSD will 454remain usable. 455.Pp 456The more unused storage there is from the SSDs point of view the 457easier a time the SSD has running its wear leveling algorithms. 458Basically the wear leveling algorithm in a modern SSD (say Intel or OCZ) 459uses a combination of static and dynamic leveling. Static is the 460best, allowing the SSD to reuse flash cells that have not been 461erased very much by moving static (unchanging) data out of them and 462into other cells that have more wear. Dynamic wear leveling involves 463writing data to available flash cells and then marking the cells containing 464the previous copy of the data as being free/reusable. Dynamic wear leveling 465is the worst kind but the easiest to implement. Modern SSDs use a combination 466of both algorithms plus also do write-combining. 467.Pp 468USB sticks often use only dynamic wear leveling and have short life spans 469because of that. 470.Pp 471In anycase, any unused space in the SSD effectively makes the dynamic 472wear leveling the SSD does more efficient by giving the SSD more 'unused' 473space above and beyond the physical space it reserves beyond its stated 474storage capacity to cycle data through, so the SSD lasts longer in theory. 475.Pp 476Write-combining is a feature whereby the SSD is able to reduced write 477amplification effects by combining OS writes of smaller, discrete, 478non-contiguous logical sectors into a single contiguous 128KB physical 479flash block. 480.Pp 481On the flip side write-combining also results in more complex lookup tables 482which can become fragmented over time and reduce the SSDs read performance. 483Fragmentation can also occur when write-combined blocks are rewritten 484piecemeal. 485Modern SSDs can regain the lost performance by de-combining previously 486write-combined areas as part of their static wear leveling algorithm, but 487at the cost of extra write/erase cycles which slightly increase write 488amplification effects. 489Operating systems can also help maintain the SSDs performance by utilizing 490larger blocks. 491Write-combining results in a net-reduction 492of write-amplification effects but due to having to de-combine later and 493other fragmentary effects it isn't 100%. 494From testing with Intel devices write-amplification can be well controlled 495in the 2x-4x range with the OS doing 16K writes, versus a worst-case 4968x write-amplification with 16K blocks, 32x with 4K blocks, and a truly 497horrid worst-case with 512 byte blocks. 498.Pp 499The 500.Dx 501.Nm 502feature utilizes 64K-128K writes and is specifically designed to minimize 503write amplification and write-combining stresses. 504In terms of placing an actual filesystem on the SSD, the 505.Dx 506.Xr hammer 8 507filesystem utilizes 16K blocks and is well behaved as long as you limit 508reblocking operations. 509For UFS you should create the filesystem with at least a 4K fragment 510size, versus the default 2K. 511Modern Windows filesystems use 4K clusters but it is unclear how SSD-friendly 512NTFS is. 513.Sh EXPLANATION OF FLASH CHIP FEATURE SIZE VS ERASE/REWRITE CYCLE DURABILITY 514Manufacturers continue to produce flash chips with smaller feature sizes. 515Smaller flash cells means reduced erase/rewrite cycle durability which in 516turn reduces the durability of the SSD. 517.Pp 518The older 34nm flash typically had a 10,000 cell durability while the newer 51925nm flash is closer to 1000. The newer flash uses larger ECCs and more 520sensitive voltage comparators on-chip to increase the durability closer to 5213000 cycles. Generally speaking you should assume a durability of around 5221/3 for the same storage capacity using the new chips versus the older 523chips. If you can squeeze out a 400TB durability from an older 40GB X25-V 524using 34nm technology then you should assume around a 400TB durability from 525a newer 120GB 310 series SSD using 25nm technology. 526.Sh WARNINGS 527I am going to repeat and expand a bit on SSD wear. 528Wear on SSDs is a function of the write durability of the cells, 529whether the SSD implements static or dynamic wear leveling (or both), 530write amplification effects when the OS does not issue write-aligned 128KB 531ops or when the SSD is unable to write-combine adjacent logical sectors, 532or if the SSD has a poor write-combining algorithm for non-adjacent sectors. 533In addition some additional erase/rewrite activity occurs from cleanup 534operations the SSD performs as part of its static wear leveling algorithms 535and its write-decombining algorithms (necessary to maintain performance over 536time). MLC flash uses 128KB physical write/erase blocks while SLC flash 537typically uses 64KB physical write/erase blocks. 538.Pp 539The algorithms the SSD implements in its firmware are probably the most 540important part of the device and a major differentiator between e.g. SATA 541and USB-based SSDs. SATA form factor drives will universally be far superior 542to USB storage sticks. 543SSDs can also have wildly different wearout rates and wildly different 544performance curves over time. 545For example the performance of a SSD which does not implement 546write-decombining can seriously degrade over time as its lookup 547tables become severely fragmented. 548For the purposes of this manual page we are primarily using Intel and OCZ 549drives when describing performance and wear issues. 550.Pp 551.Nm 552parameters should be carefully chosen to avoid early wearout. 553For example, the Intel X25V 40GB SSD has a minimum write durability 554of 40TB and an actual durability that can be quite a bit higher. 555Generally speaking, you want to select parameters that will give you 556at least 10 years of service life. 557The most important parameter to control this is 558.Va vm.swapcache.accrate . 559.Nm 560uses a very conservative 100KB/sec default but even a small X25V 561can probably handle 300KB/sec of continuous writing and still last 10 years. 562.Pp 563Depending on the wear leveling algorithm the drive uses, durability 564and performance can sometimes be improved by configuring less 565space (in a manufacturer-fresh drive) than the drive's probed capacity. 566For example, by only using 32GB of a 40GB SSD. 567SSDs typically implement 10% more storage than advertised and 568use this storage to improve wear leveling. 569As cells begin to fail 570this overallotment slowly becomes part of the primary storage 571until it has been exhausted. 572After that the SSD has basically failed. 573Keep in mind that if you use a larger portion of the SSD's advertised 574storage the SSD will not know if/when you decide to use less unless 575appropriate TRIM commands are sent (if supported), or a low level 576factory erase is issued. 577.Pp 578.Nm smartctl 579(from 580.Xr dports 7 Ap s 581.Pa sysutils/smartmontools ) 582may be used to retrieve the wear indicator from the drive. 583One usually runs something like 584.Ql smartctl -d sat -a /dev/daXX 585(for AHCI/SILI/SCSI), or 586.Ql smartctl -a /dev/adXX 587for NATA. 588Some SSDs 589(particularly the Intels) will brick the SATA port when smart operations 590are done while the drive is busy with normal activity, so the tool should 591only be run when the SSD is idle. 592.Pp 593ID 232 (0xe8) in the SMART data dump indicates available reserved 594space and ID 233 (0xe9) is the wear-out meter. 595Reserved space 596typically starts at 100 and decrements to 10, after which the SSD 597is considered to operate in a degraded mode. 598The wear-out meter typically starts at 99 and decrements to 0, 599after which the SSD has failed. 600.Pp 601.Nm 602tends to use large 64KB writes and tends to cluster multiple writes 603linearly. 604The SSD is able to take significant advantage of this 605and write amplification effects are greatly reduced. 606If we take a 40GB Intel X25V as an example the vendor specifies a write 607durability of approximately 40TB, but 608.Nm 609should be able to squeeze out upwards of 200TB due the fairly optimal 610write clustering it does. 611The theoretical limit for the Intel X25V is 400TB (10,000 erase cycles 612per MLC cell, 40GB drive, with 34nm technology), but the firmware doesn't 613do perfect static wear leveling so the actual durability is less. 614In tests over several hundred days we have validated a write endurance 615greater than 200TB on the 40G Intel X25V using 616.Nm . 617.Pp 618In contrast, filesystems directly stored on a SSD could have 619fairly severe write amplification effects and will have durabilities 620ranging closer to the vendor-specified limit. 621.Pp 622Tests have shown that power cycling (with proper shutdown) and read 623operations do not adversely effect a SSD. Writing within the wearout 624constraints provided by the vendor also does not make a powered SSD any 625less reliable over time. Time itself seems to be a factor as the SSD 626encounters defects and weak cells in the flash chips. Writes to a SSD 627will effect cold durability (a typical flash chip has 10 years of cold 628data retention when fresh and less than 1 year of cold data retention near 629the end of its wear life). Keeping a SSD cool improves its data retention. 630.Pp 631Beware the standard comparison between SLC, MLC, and TLC-based flash 632in terms of wearout and durability. Over the years, tests have shown 633that SLC is not actually any more reliable than MLC, despite having a 634significantly larger theoretical durability. Cell and chip failures seem 635to trump theoretical wear limitations in terms of device reliability. 636With that in mind, we do not recommend using SLC for anything any more. 637Instead we recommend that the flash simply be over-provisioned to provide 638the needed durability. 639This is already done in numerous NVMe solutions for the vendor to be able 640to provide certain minimum wear guarantees. 641Durability scales with the amount of flash storage (but the fab process 642typically scales the opposite... smaller feature sizes for flash cells 643greatly reduce their durability). 644When wear calculations are in years, these differences become huge, but 645often the quantity of storage needed trumps the wear life so we expect most 646people will be using MLC. 647.Pp 648Beware the huge difference between larger (e.g. 2.5") form-factor SSDs 649and smaller SSDs such as USB sticks are very small M.2 storage. Smaller 650form-factor devices have fewer flash chips and, much lower write bandwidths, 651less ram for caching and write-combining, and usb sticks in particular will 652usually have unsophisticated wear-leveling algorithms compared to a 2.5" 653SSD. It is generally not a good idea to make a USB stick your primary 654storage. Long-form-factor NGFF/M.2 devices will be better, and 2.5" 655form factor devices even better. The read-bandwidth for a SATA SSD caps 656out more quickly than the read-bandwidth for a NVMe SSD, but the larger 657form factor of a 2.5" SATA SSD will often have superior write performance 658to a NGFF NVMe device. There are 2.5" NVMe devices as well, requiring a 659special connector or PCIe adapter, which give you the best of both worlds. 660.Sh SEE ALSO 661.Xr chflags 1 , 662.Xr fstab 5 , 663.Xr disklabel64 8 , 664.Xr hammer 8 , 665.Xr swapon 8 666.Sh HISTORY 667.Nm 668first appeared in 669.Dx 2.5 . 670.Sh AUTHORS 671.An Matthew Dillon 672