Compressed Swap for Embedded Linux. Alexander Belyakov, Intel Corp.

Transcription

1 Compressed Swap for Embedded Linux Alexander Belyakov, Intel Corp.

2 Outline. 1. Motivation 2. Underlying media types 3. Related works 4. MTD compression layer driver place in kernel architecture swap-in/out data flow 5. Performance expectations 6. User-mode performance PCM NAND RAM 7. Conclusion 2

3 Motivation Memory requirements in embedded world constantly grow Most of embedded systems has both DRAM and flash memory Compress swapped pages and store them to the flash. More virtual memory available for applications More memory available for file-cache Power consumption is getting lower Entire design becomes cheaper Virtual memory subsystem becomes more healthy Performance degradation Erase cycles Wear out issues 3

4 Underlying media types 1. NAND required explicit erase bad blocks wear out/leveling issues increased software complexity 2. PCM (phase change memory) bit alterable writes cycling endurance performance 3. RAM high performance no hardware changes extra virtual memory 4

5 Related works hardware compressor/decompressor B. Tremaine, et al., IBM memory expansion technology swap caching and compression T. Cortes, et al., Improving Application Performance through Swap Compression compressed swap in RAM (embedded) Lei Yang, et al., CRAMES: Compresses RAM for Embedded Systems Compressed Caching for Linux, compressed swap on NAND (embedded) Sangduck Park, et al., Compressed Swapping for NAND Flash Memory Based Embedded Systems 5

6 MTD Compression Layer (prototype) RAM swap subsystem (VM) WRITESECT() READSECT() MTDBLOCK IF New! MTDCOMPR (ZLIB) MTD FLASH MTDCOMPR->WRITE() MTDCOMPR->READ() MTD->WRITE() MTD->READ() MTD->ERASE() MEMCPY() 6

7 MTDCOMPR swap-out data flow Write request (buffer, external offset) Fill input buffer (page size) Invalidate existing on-media data (update free space list) Compress the data Select the largest free on-media chunk from in-memory list Write compressed data (with headers) splitting between free chunks if necessary Update free space list Store external-internal offset correspondence within in-memory tree Return with success 7

8 MTDCOMPR swap-in data flow Read request (buffer, external offset) Find corresponding internal offset by external offset Read data chain from the media Decompress the data Return with success until buffer is empty 8

9 Performance expectations (flash) memory access speed RAM access speed swap access speed RAM size FLASH size Out of memory memory in use RAM only RAM + compressed SWAP on FLASH 9

10 User-mode performance (PXA271, 32MB RAM, 16MB NAND) 1000 Performance, % % Memory in use, Mbytes fixed reads random reads last reads 10

11 User-mode performance (PXA271, 32MB RAM, 16MB PCM) 1000 Performance, % % Memory in use, Mbytes fixed reads random reads last reads 11

12 Conclusion MTDCOMPR key features: increase virtual memory moderate performance impact relatively simple Media types: PCM easy to use, hardware dependent RAM easy to use, no hardware dependence NAND required complex solution, hardware dependent 32MB RAM 16MB PCM MTDCOMPR ZLIB1 80MB virtual memory perf: 100% perf: 80% 12

13 13 Backup

14 Performance expectations (prototype based data) Average performance MB RAM 32MB RAM + 16MB PCM + ZLIB Memory in use, Mbytes

15 Performance expectations (RAM) memory access speed RAM access speed swap access speed RAM disk size RAM size memory in use no swap compressed swap in RAM 15

16 Performance measurement setup System: PXA271 (416 MHz) with 32 Mbytes of RAM (Mainstone II) Linux kernel rootfs on M18 (JFFS2) Benchmarking application: Allocates memory chunk by chunk Fills it with data providing accurate compression ratio Accesses previously allocated chunks random reads fixed reads last reads Measures performance depending on amount of allocated memory Deallocates memory 16

17 User-mode performance (PXA271, 32MB RAM) 1000 Performance, % % Memory in use, Mbytes fixed reads random reads last reads 17