Integrating NAND Flash Devices onto Servers

1
Integrating NAND Flash Devices onto Servers

• By David Roberts, Taeho Kgil and David Mudge
• Presented by
• Shivananda Reddy

2
AIM

• 
  
  
  
• Computer Servers

3
Motivation

• Cost of power and cooling contributes to
  significant portion of operating cost of the data
  centre.
• System memory power (DRAM) and disk power
  contribute as mush as 50 to the total power.

4
Contributions

• Analysis of current and potential Flash usage
  models for servers.
• Argue that the extended system memory model is
  the best usage model to reduce data center
  energy.
• Two architectural modifications to improve
  NAND-based disk caches
• Splitting Flash based disk caches into read and
  write regions.
• Programmable Flash memory controller.

5
Questions ?
6
Outline

• NAND Flash overview
• NAND Flash usage models in a server
• Architecture of the Flash-based disk cache
• Architecture of the Flash memory controller
• Operation and dynamic reconfiguration
• Methodology
• Results
• Conclusion

7
NAND Flash overview

• Flash page 2KB, Flash block 128KB (64 Flash
  Pages).
• Random Flash reads and writes are performed on a
  page basis.
• Flash erasures are performed per block.
• Flash must perform an erase on a block before it
  can write to a page belonging that block.
• SLC/MLC dual mode Flash is used.

8
NAND Flash models in a server

• Extended system memory usage model A NAND Flash
  memory module is connected to the current system
  memory interface.
• Storage accelerator usage model A NAND Flash PCI
  express card is connected to the PCI express
  interface.
• Storage device usage model A Solid State Drive
  (SSD) replaces or augments the hard disk drive.
  Example SATA SSD.

9
Extended system memory

• This presents Flash as a part of the system
  memory.
• It addresses rising power contribution of DRAM.
• Electrical constraints limiting the integration
  of more system memory.
• It requires modification to the OS kernel.
• Kernel memory manager needs to be aware of unique
  organization and behavior of Flash.
• Flash reliability management can be performed by
  the kernel memory with the assistance of the
  Flash controller.

10
Storage accelerator

• This model allows the server application to
  manage Flash directly as a cache that stores
  frequently accessed code and data.
• It reduces the overall disk power by reducing the
  number of accesses to the hard disk drive.
• Flash management is distributed across the user
  application, device driver stack and the Flash
  PCI express card Firmware.

11
Storage Device

• Solid State Drive (SSD) replacing hard disk
  drive.
• Improves latency and throughput to disk and
  reduces overall disk power consumption.
• Requires modification in the kernel and a complex
  Flash device controller for Flash reliability
  management.
• A customized file system is required to fully
  exploit the benefits of Flash.
• Ties to non-Flash aware features in a hard disk
  drive interface protocol such as SATA.

12
Comparison
13
Flash-based disk cache (Arch)
14
Flash-based disk cache (Arch)

• Two level disk cache, small DRAM in front of a
  dense Flash.
• DRAM acts as a cache for the Flash.
• Flash memory controller is required for
  reliability management.
• MLC NAND Flash and is capable of switching from
  SLC mode. (Flex-OneNAND and Cho)
• Variable strength ECC is used to improve
  reliability.
• Additional data structures to manage the Flash
  blocks and pages.

15
Operating System Support

• Flash cache hash table (FCHT) stores the
  mapping between addresses on disk and Flash.
• Flash page status table (FPST) stores error
  correction code (ECC) strength, SLC/MLC mode and
  a saturating access counter. It also stores a
  valid bit field.
• Flash block status table (FBST) records the
  number of erase operations performed on each
  block as well as the degree of wear out.
• Flash global status table (FGST) tracks average
  latency and miss rate of the Flash based disk
  cache.

16
Splitting Flash into Read and Write

• Out-of-place writes degrade the performance of a
  Flash-based disk cache.
• Increases the garbage collection (GC) overhead.
• Increases the number of overall disk cache
  misses.
• Read critical Flash blocks are located in the
  read disk cache that may only evict Flash blocks
  and pages on read misses.
• Write disk cache captures all writes to the Flash
  based disk cache and performs out-of-place
  writes.
• Each read and write cache manages its own set of
  Flash blocks and LRU policy.

17
Splitting Flash into Read and Write
18
Flash memory controller
19
Flash based disk cache access
20
Methodology

• Evaluated the Flash memory controller and Flash
  device using a full system simulator called M5.
• Scaled the benchmarks, system memory size, Flash
  size and disk drive size to run on simulation
  infrastructure.
• To measure power and performance dbt2 (OLTP) and
  SPECWeb99.

21
Results

• System memory and disk energy efficiency

22
Results

• Impact of BCH code strength on system performance.

23
Results

• Flash lifetime

24
Conclusion

• Flash reduces power consumption in system
  memories and disk drives while improving the
  overall throughput.
• Integrating Flash on servers reduce the operating
  cost of a server platform.
• Proposed architecture with reliability support
  greatly improves Flash lifetime.

25
Thank You Questions ?