The disk-scheduling algorithms just discussed apply to mechanical platter- based storage like HDDs. They focus primarily on minimizing the amount of disk head movement. NVM devices do not contain moving disk heads and commonly use a simple FCFS policy. For example, the Linux NOOP scheduler uses an FCFS policy but modifies it to merge adjacent requests. The observed behavior of NVM devices indicates that the time required to service reads is uniform but that, because of the properties of flash memory, write service time is not uniform. Some SSD schedulers have exploited this property and merge only adjacent write requests, servicing all read requests in FCFS order.

Aswe have seen, I/O can occur sequentially or randomly. Sequential access is optimal for mechanical devices like HDD and tape because the data to be read or written is near the read/write head. Random-access I/O, which is measured in input/output operations per second (IOPS), causesHDDdisk head movement. Naturally, random access I/O is much faster on NVM. An HDD can produce hundreds of IOPS, while an SSD can produce hundreds of thousands of IOPS.

NVM devices offer much less of an advantage for raw sequential through- put, where HDD head seeks are minimized and reading and writing of data to the media are emphasized. In those cases, for reads, performance for the two types of devices can range from equivalent to an order of magnitude advantage for NVM devices.Writing to NVM is slower than reading, decreasing the advantage. Furthermore, while write performance for HDDs is consistent throughout the life of the device, write performance for NVM devices varies depending on how full the device is (recall the need for garbage collection and over-provisioning) and how “worn” it is. An NVM device near its end of life due to many erase cycles generally has much worse performance than a new device.

One way to improve the lifespan and performance of NVM devices over time is to have the file system inform the device when files are deleted, so that the device can erase the blocks those files were stored on. This approach is discussed further in Section 14.5.6.

Let’s lookmore closely at the impact of garbage collection on performance. Consider an NVM device under random read and write load. Assume that all blocks have beenwritten to, but there is free space available. Garbage collection must occur to reclaim space taken by invalid data. That means that a write might cause a read of one or more pages, a write of the good data in those pages to overprovisioning space, an erase of the all-invalid-data block, and the placement of that block into overprovisioning space. In summary, one write request eventually causes a page write (the data), one or more page reads (by garbage collection), and one or more page writes (of good data from the garbage-collected blocks). The creation of I/O requests not by applications but by the NVM device doing garbage collection and space management is called write amplificatio and can greatly impact the write performance of the device. In the worst case, several extra I/Os are triggered with each write request.