[B! DMA] yassã®ãƒ–ãƒƒã‚¯ãƒžãƒ¼ã‚¯

yass id:yass

DMAã«é–¢ã™ã‚‹yassã®ãƒ–ãƒƒã‚¯ãƒžãƒ¼ã‚¯ (7)

${{author_name}}$

{{{comment_expanded}}}

{{label}}

{{#is_bookmark}}ãƒªã‚¹ãƒˆ{{/is_bookmark}}{{^is_bookmark}}ãƒªãƒ³ã‚¯{{/is_bookmark}}

${{author_name}}$
{{author_name}}{{created}}
{{ #comment }}{{ comment }}{{ /comment }}
- {{ label }}

${{author_name}}$

{{{comment_expanded}}}

{{label}}

{{#is_bookmark}}ãƒªã‚¹ãƒˆ{{/is_bookmark}}{{^is_bookmark}}ãƒªãƒ³ã‚¯{{/is_bookmark}}

SSD-to-GPU Peer-to-Peer DMAã¨ãƒãƒƒãƒ•ã‚¡ç®¡ç†ï¼ˆãã®ï¼‘ï¼‰ - KaiGaiã®ä¿ºãƒ¡ãƒ¢
æ˜¨å¹´ã®æš®ã‚Œã€JPUGã‚«ãƒ³ãƒ•ã‚¡ãƒ¬ãƒ³ã‚¹ã®LTã§ã€ŽSQL+GPU+SSD=âˆžã€ã¨é¡Œã—ãŸã‚¹ãƒ”ãƒ¼ãƒã‚’è¡Œã£ãŸã€‚ SQL+GPU+SSD=âˆž (Japanese) from Kohei KaiGai www.slideshare.net ã“ã‚Œã¯ã‹ã„ã¤ã¾ã‚“ã§è¨€ãˆã°ã€ã‚¹ãƒˆãƒ¬ãƒ¼ã‚¸ã‹ã‚‰ãƒ‡ãƒ¼ã‚¿ã‚’CPU+RAMã¸ã¨ãƒãƒ¼ãƒ‰ã™ã‚‹ã‚ˆã‚Šå‰ã«ä¸€æ—¦GPUã¸ã¨ãƒ‡ãƒ¼ã‚¿ã‚’è»¢é€ã—ã€ãã“ã§ä¸è¦ãªãƒ‡ãƒ¼ã‚¿ã‚’å‰ŠãŽè½ã—ã¦ã‹ã‚‰CPU+RAMã¸ã¨æ¸¡ã—ã¦ã‚„ã‚‹äº‹ã§ã€CPUè² è·ã®è»½æ¸›ã¨RAMã®æœ‰åŠ¹æ´»ç”¨ãŒè¨ˆã‚Œã‚‹ã¨ã„ã†ã‚¢ã‚¤ãƒ‡ã‚¢ã§ã‚ã‚‹ã€‚ å®Ÿè£…ã¨ã—ã¦ã¯ã€PCI-Eãƒ‡ãƒã‚¤ã‚¹é–“ã§ã®P2P DMAæ©Ÿèƒ½ã‚’åˆ©ç”¨ã™ã‚‹äº‹ã«ã‚ˆã£ã¦NVMe SSDã®ç‰¹å®šãƒ–ãƒãƒƒã‚¯ã‹ã‚‰GPU RAMä¸Šã®ç‰¹å®šã®é ˜åŸŸã¸DMAã‚’å®Ÿè¡Œã™ã‚‹ã¨ã„ã†ã‚‚ã®ãªã®ã§ã€ã“ã“ã¯åˆ¥ã«æ–°ã—ãã‚‚ä½•ã¨ã‚‚ãªã„ã€‚ ä»¥ä¸‹ã®å›³ã¯ã€å¾“æ¥ã®ä»•çµ„ã¿ã«ãŠã‘ã‚‹ãƒ‡ãƒ¼ã‚¿ã®æµã‚Œã‚’ç¤ºã—ãŸã‚‚ã®ã€‚ SSDã‹ã‚‰èªã¿å‡ºã•ã‚ŒãŸãƒ‡ãƒ¼ã‚¿ã¯å…ˆãšCPU+RAMã«ãƒãƒƒãƒ•ã‚¡ã•ã‚Œ
yass 2016/02/14
SSD

GPU

gpgpu

dma
ãƒªãƒ³ã‚¯
e-trees.Japan, inc. :: KC705ã§PCIeã®é€šä¿¡æ€§èƒ½ã‚’æ¸¬å®šã—ã¦ã¿ãŸ
PCIe gen1 x1ã®è»¢é€é€Ÿåº¦ã¯ç†è«–(è¦æ ¼)ä¸Šã§ã¯2.5GTpsï¼Œ0.5GBpsã§ã™ï¼Ž ã—ã‹ã—ï¼Œå®Ÿéš›ã«ã¯PCIeä¸Šã®ãƒ‡ãƒ¼ã‚¿ã‚’é§†å‹•ã™ã‚‹ã‚¨ãƒ³ã‚¸ãƒ³ã®å‡¦ç†æ™‚é–“ã‚„ï¼Œãƒ‡ãƒ¼ã‚¿ä¾›çµ¦æºã‹ã‚‰ã®èªã¿å‡ºã—ï¼Œæ›¸ãè¾¼ã¿å…ˆã§ã®æ›¸ãå‡ºã—æ™‚é–“ãªã©ï¼Œã„ã‚ã„ã‚ãªã‚ªãƒ¼ãƒãƒ˜ãƒƒãƒ‰ãŒå…¥ã£ã¦ãã¾ã™ï¼Ž ã§ï¼Œå®Ÿéš›ã©ã®ãã‚‰ã„ãªã®ã‚ˆï¼Ÿã¨ã„ã†ã®ã‚’ EDKã§PCIe+DDR3ã®ã‚¢ã‚¯ã‚»ãƒ©ãƒ¬ãƒ¼ã‚¿ãƒ•ãƒ¬ãƒ¼ãƒ ãƒ¯ãƒ¼ã‚¯ã‚’ä½œã‚‹(1) ã§ä½œã£ãŸã‚·ã‚¹ãƒ†ãƒ ã§è©•ä¾¡ã—ã¦ã¿ã¾ã—ãŸï¼Ž çµæžœ ã§ï¼Œã„ããªã‚Šã§ã™ãŒï¼Œçµæžœã§ã™ï¼Ž æ¸¬å®šã—ãŸã®ã¯ï¼Œ è»¢é€é€Ÿåº¦ã¯æ¬¡ã®4ãƒ‘ã‚¿ãƒ³ï¼Œ Gen1 x1 Gen2 x1 Gen1 x4 Gen2 x4 DMAè»¢é€å˜ä½ã¯æ¬¡ã®7ãƒ‘ã‚¿ãƒ³ã‚’è©¦ã—ã¾ã—ãŸ 1KB 4KB 16KB 64KB 256KB 1MB 4MB Gen2 x4ã®å ´åˆã¯ï¼Œ DMAè»¢é€å˜ä½ãŒ1KBã®ã¨ãã«Hostâ†’DDR3/FPGAã¨DDR3/FPGAâ†’Hostã®ã©ã¡ã‚‰ã‚‚ç´„58MBps DMA
yass 2013/12/28
" Gen2 x4ã®å ´åˆã¯ï¼ŒDMAè»¢é€å˜ä½ãŒ1KBã®ã¨ãã«Hostâ†’DDR3/FPGAã¨DDR3/FPGAâ†’Hostã®ã©ã¡ã‚‰ã‚‚ç´„58MBps / DMAè»¢é€å˜ä½ãŒ4MBã®ã¨ãã«Hostâ†’DDR3/FPGAãŒç´„850MBpsï¼ŒDDR3/FPGAâ†’HostãŒç´„1035MBps / PCIeã®ç†è«–çš„ãªè»¢é€ãƒ¬ãƒ¼ãƒˆã¯Gen2 x4ã®å ´åˆ2GBps "

DMA

FPGA

PCIe

bus

bandwidth
ãƒªãƒ³ã‚¯
i350-T2 : æˆ‘æ…¢å‡ºæ¥ãšè²·ã£ã¡ã¾ã£ãŸ
æ¯”è¼ƒçš„å„ªç§€ãªLANã‚«ãƒ¼ãƒ‰è¨€ãˆã°ï¼Ÿç™ºå£²ã—ã¦ã‹ã‚‰ï¼•å¹´ä»¥ä¸ŠçµŒéŽã—ä»Šã äººæ°—ãŒè¡°ãˆãªã„Intel Gigabit CT Desktop Adapterã§ã™ãŒã€ãã‚Œã«ä»£ã‚ã‚‹ã¨æ€ã‚ã‚Œã‚‹Server Adapter i210-T1ãŒæµ·å¤–ã§ç™ºå£²ã•ã‚Œã¾ã—ãŸã€‚ä¾¡æ ¼ã¯67ãƒãƒ³ãƒ‰ç¨‹åº¦ï¼ˆæ—¥æœ¬ãƒ¬ãƒ¼ãƒˆã§ç´„10,000å††ï¼‰ã§Amazon.co.ukãªã©ã§å£²ã£ã¦ã¾ã™ã€‚æ—¥æœ¬å›½å†…ã§ã¯æœªã å£²ã£ã¦ã¾ã›ã‚“ãŒIntel.co.jpã§ã‚‚æ—¥æœ¬èªžã§å•†å“ç´¹ä»‹ãƒšãƒ¼ã‚¸ãŒæœ€è¿‘å‡ºãŸäº‹ã‚‚ã‚ã‚Šã€è¿‘ã„æ™‚æœŸï¼ˆ6æœˆãã‚‰ã„ã‹ãªï½žï¼Ÿï¼‰å£²ã‚Šå‡ºã•ã‚Œã‚‹ã¨æ€ã„ã¾ã™ã€‚ ã€ï¼–æœˆï¼’æ—¥è¿½åŠ ã€‘ã€€ã‚³ãƒ¼ãƒ‰ãƒãƒ¼ãƒ ã€ŒHaswellã€ã“ã¨ç¬¬4ä¸–ä»£Coreãƒ—ãƒã‚»ãƒƒã‚µã¨åŒæ™‚ã«å‡ºã‚‹ã¨å‹æ‰‹ã«äºˆæƒ³ã—ãŸãŒã€ã¾ã å£²ã£ã¦ãªã„ãƒ»ãƒ»ãƒ»ã‚¤ãƒ³ãƒãƒ§ãƒ«æ—¥æœ¬æ³•äººã•ã‚“ï¼ã¾ã ãƒ¼ï¼Ÿ ã€ï¼˜æœˆ24æ—¥è¿½åŠ ã€‘ã€€äºˆæƒ³ã‚ˆã‚Šï¼’ãƒ¶æœˆé…ã‚Œã¦ã€Ž ãƒ„ã‚¯ãƒ¢ãƒ‘ã‚½ã‚³ãƒ³æœ¬åº—II ã€ã‹ã‚‰è²©å£²é–‹å§‹ãªã£ãŸã‚ˆã†ã§ã™ã€‚ä¾¡æ ¼ã¯9,540å††ã£ã¨äºˆæƒ³é€šã‚Šã®ä¾¡æ ¼ã€‚æ¬²ã—ã„ãŒ50
yass 2013/08/15
"PCI Express V2.1 (5 GT/s)ã¯1ãƒãƒ¼ãƒˆå½“ãŸã‚ŠPCI Expressã§5Gbpsé€šä¿¡ã§ããƒžãƒ¼ã‚¸åˆ†ã®ä½™è£•ãŒå¢—ãˆã¾ã—ãŸ/ DMA Coalescing (DMAC)ã‚‚ç¢ºèªã§ãã¾ã—ãŸ / RSSã‚ãƒ¥ãƒ¼ã‚‚æœ€å¤§8ã‚ãƒ¥ãƒ¼ã¾ã§è¨å®šã§ãã‚‹ã‚ˆã†ã«ãªã£ã¦ã¾ã™ãã€‚"

intel

nic

rss

i350

DMA

PCIe

bandwidth
ãƒªãƒ³ã‚¯
ã€ç¬ åŽŸä¸€è¼ã®ãƒãƒ¼ãƒ‰ã‚¦ã‚§ã‚¢ãƒˆãƒ¬ãƒ³ãƒ‰ã€‘ ã‚¢ãƒ³ã‚³ã‚¢ã‚’å¤§å¹…ã«é€²åŒ–ã•ã›ã€å‡¦ç†èƒ½åŠ›ã¨é›»åŠ›åŠ¹çŽ‡ã‚’æ”¹å–„ã—ãŸXeon E5 2600ã‚·ãƒªãƒ¼ã‚º
yass 2013/07/03
"Intelã¯DDIOï¼ˆData Direct I/Oï¼‰ã¨å‘¼ã°ã‚Œã‚‹ä»•çµ„ã¿ã‚’å°Žå…¥/DDR3-1600ã®ãƒ¡ãƒ¢ãƒªãƒ¢ã‚¸ãƒ¥ãƒ¼ãƒ«ã‚’åˆ©ç”¨ã—ãŸå ´åˆã€1ãƒãƒ£ãƒãƒ«ã‚ãŸã‚Š12.8G/ç§’ã®å¸¯åŸŸå¹…ã¨ãªã‚‹ãŸã‚ã€12.8GB/ç§’X4ï¼51.2GB/ç§’ã®å¸¯åŸŸå¹…"

Intel

CPU

xeon

SandyBridge

AVX

DDIO

DMA

bandwidth

ddr

memory
ãƒªãƒ³ã‚¯
ã‚³ã‚¤ãƒ³åž‹ãƒãƒƒãƒ†ãƒªãƒ¼: Dell XPS 8300 ã‚µãƒ¼ãƒ“ã‚¹ãƒžãƒ‹ãƒ¥ã‚¢ãƒ« : User Guide : Documentation
ã‚ã‚‰ã‚†ã‚‹ãƒ‹ãƒ¼ã‚ºã‚’1ã‹æ‰€ã§ã‚µãƒãƒ¼ãƒˆã€‚ã‚µã‚¤ãƒ³ã‚¤ãƒ³ã™ã‚‹ã¨ã€ãƒ‘ãƒ¼ã‚½ãƒŠãƒ©ã‚¤ã‚ºã•ã‚ŒãŸãƒ˜ãƒ«ãƒ—ã‚’å—ã‘ãŸã‚Šã€ç™»éŒ²æ¸ˆã¿ã®ãƒ‡ãƒã‚¤ã‚¹ã€ã‚½ãƒ•ãƒˆã‚¦ã‚§ã‚¢ã€ãŠã‚ˆã³æ—¢å˜ã®ã‚µãƒ¼ãƒ“ã‚¹ ãƒªã‚¯ã‚¨ã‚¹ãƒˆã«ã‚¢ã‚¯ã‚»ã‚¹ã—ãŸã‚Šã§ãã¾ã™ã€‚
yass 2012/12/24
intel

nic

ioat

dma
ãƒªãƒ³ã‚¯
Diving into the Linux Networking Stack, Part I | Beyond Syntax | Michael J. Schultz
This is the first part of a (planned) multi-part series on receiving a network packet on a modern Linux kernel. Have you ever wondered what happens when your Linux machine gets a packet from the network? I have, but most of the information I've seen is concerned with the 2.4.x kernel. For my own sake, I decided to take a walk through the Linux networking stack (using Linux kernel 2.6.37) and thoug
yass 2012/12/23
Linux

Network

kernel

nic

cpu

interrupt

rss

napi

msi-x

dma
ãƒªãƒ³ã‚¯
X86 hardware for packet processing
This document discusses hardware and software requirements for high-performance packet processing. It covers Direct Cache Access (DCA) which allows network cards to directly write received packets to the CPU cache to reduce memory traffic. It also discusses virtualization techniques like Virtual Machine Device Queues (VMDq) and Receive-Side Scaling (RSS) which improve performance by distributing n
yass 2012/10/11
" Processing time 67 nsec for a packet â€¢ About 134 cycles for Xeon 2GHz "

CPU

hardware

network

nic

packet

multiqueue

dca

rss

pps

dma
ãƒªãƒ³ã‚¯
1