Citation Context

...ordering, such as Linux ext2, additional ordering and writes are needed. This problem of recursive update has been solved in other systems by adding a level of indirection (e.g., the inode map in LFS =-=[27]-=-). 3.2 Segmented Address Space To solve the recursive update problem without requiring substantial changes to the existing file system, we introduce a segmented address space with two segments (see Fi...

Citation Context

...on allows the system to keep active blocks in mirrored storage for performance reasons, and move inactive blocks to RAID to increase effective capacity [32] and overcome the RAID small-update problem =-=[26]-=-. When a file system runs atop a RAID, excess indirection exists because the file system maps logical offsets to logical block addresses. The RAID, in turn, maps logical block addresses to physical (d...

Citation Context

...orting physical resources. Another concern is the extra level of information naturally exported by exposing physical names to clients. Although the value of physical names has been extolled by others =-=[12]-=-, a device manufacturer may feel that such information reveals too much of their “secret sauce” and thus be wary of adopting such an interface. We believe that if such a concern exists, the device cou...

Citation Context

...higher and lower layers of the system. Without such a constraint, one could often remove the excess indirection and thus improve the system. For example, if an OS running on a para-virtualized system =-=[31]-=- is modified to request a machine page from the hypervisor and then install the correct virtual-to-machine page translation in its page tables, the hypervisor is relieved of having to manage this extr...

Citation Context

... atop a RAID array, and the focus of our work, file systems atop flash-based SSDs. An excellent example of excess indirection arises in memory management of operating systems running atop hypervisors =-=[9]-=-. The OS manages virtual-to-physical mappings for each process that is running; the hypervisor, in turn, manages physical-to-machine mappings for each OS. In this manner, the hypervisor has full contr...

Citation Context

...space and time overheads in virtualized systems. The space overhead comes from maintaining OS physical addresses to machine addresses mapping for each page and from possible additional space overhead =-=[1]-=-. Time overheads exist as well in cases like the MIPS TLB-miss lookup in Disco [9]. Indirection also exists in modern disks. For example, modern disks maintain a small amount of extra indirection that...

Citation Context

...s Solid State Devices (SSDs). In modern SSDs, an indirection map in the Flash Translation Layer (FTL) enables the device to map writes in its virtual address space to any underlying physical location =-=[11, 14, 16, 19, 21, 22]-=-. FTLs use indirection for two reasons: first, to transform the erase/program cycle mandated by flash into the more typical write-based interface via copy-on-write techniques, and second, to implement...

Citation Context

...mapping and hybrid mapping FTLs, and a wear-leveling algorithm on all three FTLs that considers a block’s remaining erase cycles and its data temperature similar to a previous wear-leveling algorithm =-=[2]-=-. System Setup: We implemented the emulated nameless-write device and the nameless-writing ext3 file system on a 64-bit Linux 2.6.33 kernel. The page-level Image Size Page Hybrid Nameless 328 MB 328 K...

Citation Context

...rection for two reasons: first, to transform the erase/program cycle mandated by flash into the more typical write-based interface via copy-on-write techniques, and second, to implement wear leveling =-=[18, 20]-=-, which is critical to increasing SSD lifetime. Because a flash block becomes unusable after a certain number of erase-program cycles (10,000 or 100,000 cycles according to manufacturers [8, 15]), suc...

Citation Context

...h. With more inconsistent metadata, the overhead of recovery is higher. 7 Related Work A large body of work on flash-based SSD FTLs and file systems that manage them has been proposed in recent years =-=[11, 14, 16, 19, 21, 22, 25, 33]-=-. In this section, we discuss the two research projects that are most related to nameless writes. Range writes [5] use an approach similar to nameless writes. Range writes were proposed to improve har...

Citation Context

...al metadata writes, we propose in-place garbage collection, which writes the live data back to the same location instead of migrating it. A similar hole-plugging approach was proposed in earlier work =-=[24]-=-, where live data is used to plug the holes of most utilized segments. To perform in-place garbage collection, the FTL selects a candidate block using a certain policy. The FTL reads all live pages fr...

Citation Context

...s Solid State Devices (SSDs). In modern SSDs, an indirection map in the Flash Translation Layer (FTL) enables the device to map writes in its virtual address space to any underlying physical location =-=[11, 14, 16, 19, 21, 22]-=-. FTLs use indirection for two reasons: first, to transform the erase/program cycle mandated by flash into the more typical write-based interface via copy-on-write techniques, and second, to implement...

Citation Context

...s Solid State Devices (SSDs). In modern SSDs, an indirection map in the Flash Translation Layer (FTL) enables the device to map writes in its virtual address space to any underlying physical location =-=[11, 14, 16, 19, 21, 22]-=-. FTLs use indirection for two reasons: first, to transform the erase/program cycle mandated by flash into the more typical write-based interface via copy-on-write techniques, and second, to implement...

Citation Context

...es include ideas for “smart” disks that remap writes in order to improve performance (for example, by writing to the nearest free location), which have been explored in previous research such as Loge =-=[13]-=- and “intelligent” disks [30]. These smart disks require large indirection tables inside the drive to map the logical address of the write to its current physical location. This requirement introduces...

Citation Context

...s Solid State Devices (SSDs). In modern SSDs, an indirection map in the Flash Translation Layer (FTL) enables the device to map writes in its virtual address space to any underlying physical location =-=[11, 14, 16, 19, 21, 22]-=-. FTLs use indirection for two reasons: first, to transform the erase/program cycle mandated by flash into the more typical write-based interface via copy-on-write techniques, and second, to implement...

Citation Context

...s Solid State Devices (SSDs). In modern SSDs, an indirection map in the Flash Translation Layer (FTL) enables the device to map writes in its virtual address space to any underlying physical location =-=[11, 14, 16, 19, 21, 22]-=-. FTLs use indirection for two reasons: first, to transform the erase/program cycle mandated by flash into the more typical write-based interface via copy-on-write techniques, and second, to implement...

Citation Context

...ling [18, 20], which is critical to increasing SSD lifetime. Because a flash block becomes unusable after a certain number of erase-program cycles (10,000 or 100,000 cycles according to manufacturers =-=[8, 15]-=-), such indirection is needed to spread the write load across flash blocks evenly and thus ensure that no particularly popular block causes the device to fail prematurely. Unfortunately, the indirecti...

Citation Context

...acts the needed values from the lower level indirection and installs them directly. De-indirection has been successfully applied in a few domains, most notably within hypervisors. The Turtles project =-=[7]-=- provides an excellent example: in a recursively-virtualized environment (with hypervisors running on hypervisors), the Turtles system installs what the authors refer to as multi-dimensional page tabl...

Citation Context

...ling [18, 20], which is critical to increasing SSD lifetime. Because a flash block becomes unusable after a certain number of erase-program cycles (10,000 or 100,000 cycles according to manufacturers =-=[8, 15]-=-), such indirection is needed to spread the write load across flash blocks evenly and thus ensure that no particularly popular block causes the device to fail prematurely. Unfortunately, the indirecti...

Citation Context

...to the virtual block space in our nameless-writing implementation, the mapping table size of the nameless-writing device is dependent on the metadata size of the file system image. We use Impressions =-=[3]-=- to create typical file system images of sizes up to 1 TB and calculate their metadata sizes. Figure 3 shows the mapping table sizes of the three FTLs with different file system images produced by Imp...

Citation Context

...s Solid State Devices (SSDs). In modern SSDs, an indirection map in the Flash Translation Layer (FTL) enables the device to map writes in its virtual address space to any underlying physical location =-=[11, 14, 16, 19, 21, 22]-=-. FTLs use indirection for two reasons: first, to transform the erase/program cycle mandated by flash into the more typical write-based interface via copy-on-write techniques, and second, to implement...

Citation Context

...nts to a data block by the inode number, the inode generation number, and the offset of the block in the inode. For file systems that already explicitly record back references, such as btrfs and NoFS =-=[10]-=-, the back references can simply be reused for our purposes. Such metadata structure identification can be used in several tasks. First, when searching for a data block in the page cache, we obtain th...

Citation Context

...rection for two reasons: first, to transform the erase/program cycle mandated by flash into the more typical write-based interface via copy-on-write techniques, and second, to implement wear leveling =-=[18, 20]-=-, which is critical to increasing SSD lifetime. Because a flash block becomes unusable after a certain number of erase-program cycles (10,000 or 100,000 cycles according to manufacturers [8, 15]), suc...

Citation Context

...disks that remap writes in order to improve performance (for example, by writing to the nearest free location), which have been explored in previous research such as Loge [13] and “intelligent” disks =-=[30]-=-. These smart disks require large indirection tables inside the drive to map the logical address of the write to its current physical location. This requirement introduces new reliability challenges, ...

Citation Context

...tures can point to the same data block, which may result in a large amount of associated metadata. We can use file system intrinsic back references, such as those in btrfs, or structures like Backlog =-=[23]-=- to represent associated metadata. Another problem is that multiple metadata blocks need to be updated after a nameless write. One possible way to control the number of metadata updates is to add a li...

Citation Context

...that doing so both reduces space and time overheads, thus making for simpler, less costly, and higher-performance SSD-based storage. 1 Introduction Indirection is a core technique in computer systems =-=[28]-=-. Whether in the mapping of file names to blocks, or a virtual address space to an underlying physical one, system designers have applied indirection to improve system performance, reliability, and ca...

Citation Context

...stems that manage them has been proposed in recent years [11, 14, 16, 19, 21, 22, 25, 33]. In this section, we discuss the two research projects that are most related to nameless writes. Range writes =-=[5]-=- use an approach similar to nameless writes. Range writes were proposed to improve hard disk performance by letting the file system specify a range of addresses and letting the device pick the final p...

Citation Context

...h. With more inconsistent metadata, the overhead of recovery is higher. 7 Related Work A large body of work on flash-based SSD FTLs and file systems that manage them has been proposed in recent years =-=[11, 14, 16, 19, 21, 22, 25, 33]-=-. In this section, we discuss the two research projects that are most related to nameless writes. Range writes [5] use an approach similar to nameless writes. Range writes were proposed to improve har...

Citation Context

...w present the overall application performance of nameless-writing, page-level mapping and hybrid mapping FTLs with macro-benchmarks. We use varmail, fileserver, and webserver from the filebench suite =-=[29]-=-. Figure 2 shows the throughput of these benchmarks. We see that both page-level mapping and namelesswriting FTLs perform better than the hybrid mapping FTL with varmail and fileserver. These benchmar...

Citation Context

...sh recovery correct and what are its overheads? SSD Emulator: We built an SSD emulator which models a multi-plane SSD with garbage collection and wear leveling as a pseudo block device based on David =-=[4]-=-. We implemented three types of FTLs: page-level mapping, hybrid mapping and nameless-writing on top of the PSU objected-oriented SSD simulator codebase [6]. Data is stored in memory to enable quick a...