Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
Mackerelにおける時系列データベースの性能改善 / Performance Improv...
Search
Sponsored
·
Ship Features Fearlessly
Turn features on and off without deploys. Used by thousands of Ruby developers.
→
Yuuki Tsubouchi (yuuk1)
July 09, 2016
Technology
9k
13
Share
Mackerelにおける時系列データベースの性能改善 / Performance Improvement of TSDB in Mackerel
ペパボ・はてな技術大会〜インフラ技術基盤〜@福岡
Yuuki Tsubouchi (yuuk1)
July 09, 2016
More Decks by Yuuki Tsubouchi (yuuk1)
See All by Yuuki Tsubouchi (yuuk1)
AIスーパーコンピュータにおけるLLM学習処理性能の計測と可観測性 / AI Supercomputer LLM Benchmarking and Observability
yuukit
1
860
SREはサイバネティクスの夢をみるか? / Do SREs Dream of Cybernetics?
yuukit
3
490
SREのためのテレメトリー技術の探究 / Telemetry for SRE
yuukit
13
3.6k
AIスパコン「さくらONE」の オブザーバビリティ / Observability for AI Supercomputer SAKURAONE
yuukit
2
1.4k
AIスパコン「さくらONE」のLLM学習ベンチマークによる性能評価 / SAKURAONE LLM Training Benchmarking
yuukit
2
1.1k
とあるSREの博士「過程」 / A Certain SRE’s Ph.D. Journey
yuukit
11
6.9k
eBPFを用いたAIネットワーク監視システム論文の実装 / eBPF Japan Meetup #4
yuukit
3
1.7k
クラウドのテレメトリーシステム研究動向2025年
yuukit
4
1.2k
博士論文公聴会: Scaling Telemetry Workloads in Cloud Applications: Techniques for Instrumentation, Storage, and Mining / PhD Defence
yuukit
1
710
Other Decks in Technology
See All in Technology
生成AIはソフトウェア開発の革命か、ソフトウェア工学の宿題再提出なのか -ソフトウェア品質特性の追加提案-
kyonmm
PRO
2
880
PdM・Eng・QAで進めるAI駆動開発の現在地/aidd-with-pdm-eng-qa
shota_kusaba
0
200
セキュリティ対策、何からはじめる? CloudNative環境の脅威モデリングと リスク評価実践入門 #cloudnativekaigi
varu3
5
790
「背中を見て育て」からの卒業 〜専門技術としてのテスト設計を軸に、品質保証のバトンを繋ぐ〜 #genda_tech_talk
nihonbuson
PRO
3
1.2k
そのSLO 99.9%、本当に必要ですか? 〜優先度付きSLOによる責任共有の設計思想〜 / Is that 99.9% SLO really necessary? Design philosophy of shared responsibility through prioritized SLOs
vtryo
0
580
大学職員のための生成AI最前線 :最前線を、AIガバナンスとして読み直すためのTips
gmoriki
2
4k
カオナビに Suspenseを導入するまで / The Road to Suspense at kaonavi
kaonavi
1
450
Every Conversation Counts
kawaguti
PRO
0
210
変化の激しい時代をゴキゲンに生き抜くために 〜ストレスマネジメントのススメ〜
kakehashi
PRO
5
1.3k
Forget technical debt
ufried
0
180
ボトムアップの改善の火を灯し続けろ!〜支援現場で学んだ、消えないための3つの打ち手〜 / 20260509 Kazuki Mori
shift_evolve
PRO
2
650
20260513_生成AIを専属DSに_AI分析結果の検品テクニック_ハンズオン_交通事故データ
doradora09
PRO
0
220
Featured
See All Featured
Keith and Marios Guide to Fast Websites
keithpitt
413
23k
Prompt Engineering for Job Search
mfonobong
0
300
The Cost Of JavaScript in 2023
addyosmani
55
9.9k
Winning Ecommerce Organic Search in an AI Era - #searchnstuff2025
aleyda
1
2k
SEO Brein meetup: CTRL+C is not how to scale international SEO
lindahogenes
1
2.6k
Unsuck your backbone
ammeep
672
58k
Fashionably flexible responsive web design (full day workshop)
malarkey
408
66k
Making the Leap to Tech Lead
cromwellryan
135
9.8k
Redefining SEO in the New Era of Traffic Generation
szymonslowik
1
300
Navigating Algorithm Shifts & AI Overviews - #SMXNext
aleyda
1
1.2k
The AI Search Optimization Roadmap by Aleyda Solis
aleyda
1
5.8k
Documentation Writing (for coders)
carmenintech
77
5.3k
Transcript
Mackerelʹ͓͚Δ ࣌ܥྻσʔλϕʔεͷੑೳվળ ϖύϘɾͯͳٕज़େձʙΠϯϑϥٕज़ج൫ʙ@Ԭ ͯͳ id:y_uuki
id:y_uuki yuuki ΣϒΦϖϨʔγϣϯΤϯδχΞ@ͯͳ ೖࣾ3͘Β͍
07/02@ژ https://speakerdeck.com/yuukit/linux-network-performance-improvement-at-hatena
͘͡ 1. Mackerelͱ࣌ܥྻσʔλ 2. GraphiteͷΞʔΩςΫνϟͱੑೳঢ়گ 3. σΟεΫεϥογϯάͱͦͷղܾ 4. ·ͱΊ
͘͡ 1. Mackerelͱ࣌ܥྻσʔλ 2. GraphiteͷΞʔΩςΫνϟͱੑೳঢ়گ 3. σΟεΫεϥογϯάͱͦͷղܾ 4. ·ͱΊ
https://mackerel.io
αʔόͷϝτϦοΫՄࢹԽ
MackerelͷΞʔΩςΫνϟ
Mackerelͷ࣌ܥྻσʔλͷಛੑ • ΤʔδΣϯτ͕Ϣʔβ͞Μͷϗετ͔ΒຖϝτϦοΫ ߘ • 2016/01࣌ͰΞΫςΟϒΤʔδΣϯτ 10,000+ • 1ΤʔδΣϯτ͋ͨΓͷϝτϦοΫ࠷େ200 •
ԾʹฏۉϝτϦοΫΛ100 metrics/agentͱ͢Δͱɹ ߹ܭૹ৴ϝτϦοΫ 1,000,000 metrics/min + • ϝτϦοΫͷେྔॻ͖ࠐΈʹ͑ΒΕΔσʔλϕʔε͕ ඞཁ
Graphite
͘͡ 1. Mackerelͱ࣌ܥྻσʔλ 2. GraphiteͷΞʔΩςΫνϟͱੑೳঢ়گ 3. σΟεΫεϥογϯάͱͦͷղܾ 4. ·ͱΊ
Graphiteͱ • PythonͰॻ͔Εͨ࣌ܥྻσʔλϕʔεϛυϧΣΞ • HTTPΠϯλϑΣʔε ʢॻ͖ࠐΈಠࣗϓϩτίϧʣ • ग़ྗσʔλܗࣜάϥϑը૾·ͨJSON Graphite (timestamp,
name, value) graph request Image or JSON
GraphiteͷΞʔΩςΫνϟ (timestamp, name, value) graph request Image or JSON carbon
graphite-web filesystem write read whisper whisper
GraphiteͷΞʔΩςΫνϟ (graphite-web) (timestamp, name, value) graph request Image or JSON
carbon graphite-web filesystem write read whisper whisper ಡΈࠐΈཁٻΛड͚͚ΔͨΊͷWebΞϓϦέʔγϣϯ
GraphiteͷΞʔΩςΫνϟ (carbon) (timestamp, name, value) graph request Image or JSON
carbon graphite-web filesystem write read whisper whisper ॻ͖ࠐΈཁٻΛड͚͚ΔͨΊͷσʔϞϯ
GraphiteͷΞʔΩςΫνϟ (whisper) (timestamp, name, value) graph request Image or JSON
carbon graphite-web filesystem write read whisper whisper ࣌ܥྻDBϑΝΠϧΛ࡞ɾߋ৽͢ΔͨΊͷϥΠϒϥϦ ϝτϦοΫ͝ͱʹ ϑΝΠϧ͕Ͱ͖Δ
Whisperͷσʔλߏ • ͯ͢ͷσʔλΛอଘ͢ΔͱσΟεΫ༻ྔ͕ංେԽ • timestamp: 4byte, value: 8byteͱͯ͠12bytes/datapointͱ͢Δ ͱɺ1Ͱ6MB/metric •
ݹ͍σʔλʹ͍ͭͯҰఆظؒͰฏۉԽor࠷େΛؙͯ͠Ί ͯ͠·ͬͯσΟεΫ༻ྔΛઅ • ex. 1ਫ਼ͷσʔλ1͚ͩͰΑ͍͕ɺ5ਫ਼ͷσʔλ 1िؒ͢ͱ͍͏Α͏ͳΠϝʔδ
Graphiteͷॻ͖ࠐΈύϑΥʔϚϯεಛੑ(CPUར༻) • carbon2ͭͷεϨου͕ڠௐͯ͠ಈ࡞͢Δ • σʔλΛड͚औΔωοτϫʔΫI/OεϨου • ϑΝΠϧॻ͖ࠐΈͷͨΊͷI/OεϨου • ΠϕϯτۦಈϞσϧͷωοτϫʔΫαʔό •
όοϑΝ͝͠ʹεϨουؒͰσʔλϙΠϯτΛ͢ • ֤εϨου͕1ίΞͰ͢Δ • carbonϓϩηεΛෳݸͨͯͯࢄͤ͞Δ
Graphiteͷॻ͖ࠐΈύϑΥʔϚϯεಛੑ(σΟεΫIO) • େྔͷϑΝΠϧʹখ͞ͳσʔλྔʢ12ByteʣΛ1Ҏ ʹॻ͖ࠐΉ • ϑΝΠϧγεςϜ্ͷۙྡϒϩοΫʹ·ͱΊͯॻ͘͜ͱ ͕Ͱ͖ͳ͍ͨΊɺI/Oޮѱ͍ (શํҐॻ͖ࠐΈ) • ໘ɺಉ࣌ʹෳͷεϨου͕1ͭͷϑΝΠϧʹॻ͖ࠐ
Ή͜ͱ͕ͳ͍ͨΊɺ I/OͷฒྻߴΊ͍͢ • XFSͷΑ͏ͳฒྻI/Oʹ༏ΕͨϑΝΠϧγεςϜͰͳ͘ ͯɺੑೳมΘΒͳ͍ (ext4ͳͲ)
ϋʔυΣΞߏͱϦιʔε༻ྔ • CPU: Xeon E5-2697 v3 @ 2.60GHz 2 socket
28ίΞ • ϝϞϦ: 126GB • σΟεΫ: Fusion ioMemory ioDrive2 6.4TB • ͍ΘΏΔϑϨογϡετϨʔδɻϝʔΧʔެশ 300k write IOPS • ࣮ޮI/Oੑೳ: 50k ~ 100k write IOPS • ී௨ͷSSDͳΒ1/10ͷੑೳ͕ͰΕྑ͍ํ
Graphiteνϡʔχϯά • ioDriveͷIOPSΛ͍ΔલʹCPUϦιʔεΛ͍͖ͬ ͯ͠·͏ͨΊɺCPUΛઅͯ͠I/Oʹ͚Δߟ͑ํ • random writeʹڧ͍ߴͳσΟεΫͳͨΊɺجຊతʹ carbonI/Oεέδϡʔϥʹ༨ܭͳ࠷దԽΛͤ͞ͳ͍ • ιʔτʹΑΔI/OޮԽI/OϦιʔεΛ͍͖Βͳ͍
ͨΊͷ੍ݶͷύϥϝʔλ͕͋Δ • echo noop > /sys/block/fioa/queue/scheduler
GraphiteΫϥελߏ (timestamp, name, value) graphite-web carbon carbon … … LB
carbon carbon … … LB LB carbon carbon … …
ৄ͘͠ϒϩάͰ http://blog.yuuk.io/entry/high-performance-graphite
͘͡ 1. Mackerelͱ࣌ܥྻσʔλ 2. GraphiteͷΞʔΩςΫνϟͱੑೳঢ়گ 3. σΟεΫεϥογϯάͱͦͷղܾ 4. ·ͱΊ
write IOPS read IOPS ಥવͷreadෛՙ૿େ
ͳʹ͕ى͖ͨͷ͔ • read IOPS͕૿Ճ͠ɺwrite IOPS͕ݮগ͍ͯ͘͠ • ϝϞϦෆʹΑΔSwapྖҬͷ༻ͳ͠ɻOSͷϝϞϦ ༻ྔ1/3ఔͩͬͨ • αʔϏεͷಥൃతͳΞΫηε૿Ճͳ͠
• sar -BͰɺҰఆ࣌ؒͷϖʔδΠϯͱϖʔδΞτͷ͕ ҟৗʹ૿͍͑ͯͨ͜ͱ͕໌ • ͜ͷݱΛσΟεΫεϥογϯάͱݺͿ͜ͱʹ͢Δ • LinuxͷϖʔδΩϟογϡͷΈͱGraphiteͷI/Oύ λʔϯ͔ΒݪҼΛਪͨ͠
LinuxͷϖʔδΩϟογϡ • ϝϞϦͷ༁ = used + buffers/caches + free •
ϑΝΠϧγεςϜ͔ΒσʔλΛಡΈࠐΉ/ॻ͖ࠐΉͱɺ࣍ճ Ҏ߱ߴʹಡ·ͤΔͨΊʹɺOS͕ϖʔδ୯ҐͰσΟεΫ্ ͷσʔλΛϝϞϦʹࡌͤΔ • ϖʔδΩϟογϡͱݺͿ • ϖʔδΩϟογϡLRUΞϧΰϦζϜɻ࠷ۙࢀর͞Εͨ Ωϟογϡσʔλ͠ɺࢀর͞Εͳ͍ݹ͍Ωϟογϡσʔ λΛফ͢ • ϖʔδΩϟογϡ௨ৗϝϞϦ༻ྔʹؚ·Εͳ͍
GraphiteͷI/Oύλʔϯ • 1ҎʹશͯͷΞΫςΟϒͳwhisperϑΝΠϧʹॻ͖ ࠐΉͨΊɺσΟεΫͷൣғʹͬͯॻ͖ࠐΈ͕Δ • whisperͷϝτϦοΫॻ͖ࠐΈૢ࡞ɺwrite(2)͚ͩͰ ͳ͘ɺϝλσʔλͷಡΈࠐΈΦϑηοτܭࢉͷͨΊ ͷread(2)Δ • ϖʔδΩϟογϡread͚ͩͰͳ͘writeʹ༗ޮ
(Direct I/Oআ͘) • GraphiteϗετେྔͷϖʔδΩϟογϡΛͭ
read IOPS૿ͷݪҼ • ϖʔδΠϯͱϖʔδΞτճ͕ଟ͍ͱ͍͏͜ͱɺ LRUʹΑΓݹ͍Ωϟογϡ͕͍ग़͞Ε͍ͯΔ • whisperॻ͖ࠐΈͷreadͰϖʔδΩϟογϡ͕ޮ͔ͳ͘ ͳͬͨ݁Ռɺread IOPS͕૿͑ͨ Memory
used page cache page in page out
ϖʔδΩϟογϡͷઅ • ࡌϝϞϦΛ૿͢͜ͱͰҰԠղܾͰ͖Δ͕ɺ͢Ͱʹ 126GB RAMͳͷͰɺແବͳϖʔδΩϟογϡΛݮ͍ͨ͠ • writeͨ͠σʔλΛ͙͢ʹಡΉͱݶΒͳ͍ͨΊɺwrite࣌ͷ σʔλΛΩϟογϡʹͷͤͳ͍ => Direct
I/O • ͔͠͠ɺDirect I/OΛ͏ͨΊʹɺϒϩοΫαΠζͰϝϞ ϦΞϥΠϝϯτΛἧ͑Δඞཁ͕͋Δ => PythonͰΔͷ͕ ͱͯ໘ (malloc => posix_memalign) • posix_fadvise(2)Λͬͯղܾ
posix_fadvise(2) • ϓϩηε͕ΧʔωϧϑΝΠϧσʔλͷΞΫηεύλʔϯΛ ௨ • Χʔωϧࢦఆ͞ΕͨΞΫηεύλʔϯʹԠͯ͡I/Oੑೳ͕ ্͢ΔΑ͏ʹ࠷దԽ • ΞΫηεύλʔϯ •
POSIX_FADV_SEQUENTIAL: 2ഒͷઌಡΈ • POSIX_FADV_RANDOM: ઌಡΈఀࢭ • POSIX_FADV_DONTNEED: Ωϟογϡͨ͠ϖʔδͷղ์ • etc int posix_fadvise(int fd, off_t offset, off_t len, int advice);
posix_fadvise(2)ΛGraphiteʹద༻ • ࠷ॳɺϖʔδΩϟογϡΛམͱ͢Φϓγϣϯʹண • whisperͷॻ͖ࠐΈϩδοΫ݁ߏෳࡶͳͨΊɺwriteʹ ΑΔϖʔδΩϟογϡ෦͚ͩΛམͱ͢ͷ͕͍͠ • FAD_RANDONʹΑΓɺઌಡΈΛͤͣඞཁͳϖʔδ͚ͩ Ωϟογϡ͢ΔΑ͏ʹͨ͠ •
whisperͷॻ͖ࠐΈͰγʔέϯγϟϧʹᢞΊΔॲཧͳ͍ • ઌಡΈ͍ͯͨ͠ແବͳϖʔδΩϟογϡ͕ݮͬͨ Active(file): 5387160 kB Inactive(file): 37566804 kB Active(file): 32252136 kB Inactive(file): 7231020 kB /proc/meminfo before & after
GraphiteͷPull Request
Pull Request༰ • มߋ༰͞΄Ͳ͘͠ͳ͍ • fadvise ϞδϡʔϧΛ͏ • straceͯ͠posix_fadvise͕Ͱͯ͘Εok •
ৗʹfadvise͢Δͷ͕Α͍͔Θ͔Βͳ͍ͨΊɺઃఆϑΝΠϧ ʹΑΔ༗ޮɾແޮΛΓସ͑ΒΕΔΑ͏ʹ (σϑΥϧτແޮ) • Ϛʔδͯ͠Β͏·Ͱ1ϲ݄͘Β͍͔͔ͬͨ with open(path, 'r+b') as fh: if CAN_FADVISE and FADVISE_RANDOM: posix_fadvise(fh.fileno(), 0, 0, POSIX_FADV_RANDOM)
ςετεΫϦϓτʹΑΔݕূ https://gist.github.com/yuuki/8d5d386115b0f01b5371 • whisperͷॻ͖ࠐΈؔΛͬͯɺ࣮ࡍʹϖʔδΩϟο γϡͷྔ͕ݮΔ͔Ͳ͏͔֬ೝ • 100ݸͷwhisperϑΝΠϧʹରͯ͠100ݸͷσʔλϙΠϯ τΛॻ͖ࠐΉεΫϦϓτ • /proc/<pid>/io
ͷread_bytes(࣮ࡍʹσΟεΫ͔ΒಡΈͩ ͨ͠αΠζ)ΛΈΔ • POSIX_FAD_RANDOMΦϓγϣϯΛ͚ͭΔͱϖʔδ Ωϟογϡྔ͕1/2ʹͳͬͨ
͘͡ 1. Mackerelͱ࣌ܥྻσʔλ 2. GraphiteͷΞʔΩςΫνϟͱੑೳঢ়گ 3. σΟεΫεϥογϯάͱͦͷղܾ 4. ·ͱΊ
·ͱΊ • MackerelͰ 1,000,000 metrics/min + ͷϝτϦοΫ ॻ͖ࠐΈΛࡹ͘ඞཁ͕͋Δ • ࣌ܥྻσʔλϕʔεͱͯ͠GraphiteΛબ
• ioDriveલఏͰOSͤͷνϡʔχϯά • σΟεΫεϥογϯάΛposix_fadviseʹΑΓ writebackʹΑΔϖʔδΩϟογϡΛແޮʹ͢Δύον Ͱղܾ
None
1ҎԼͷཻͷϝτϦοΫ ཻΛଛͳΘͣظอଘ ϦΞϧλΠϜͳҟৗݕ
࣍ੈͷ࣌ܥྻσʔλϕʔεʹ ৽͍ͨ͠
http://hatenacorp.jp/recruit/fresh/operation-engineer ٕज़͕͖ͳਓ
ຊεϥΠυͷKeynoteςϯϓϨʔτͱͯ͠ shoya140͞ΜͷZebra(http://shoya.io/blog/zebra/) ΛΘ͍͖ͤͯͨͩ·ͨ͠ Mackerelʹ͓͚Δ ࣌ܥྻσʔλϕʔεͷੑೳվળ ϖύϘɾͯͳٕज़େձʙΠϯϑϥٕज़ج൫ʙ@Ԭ ͯͳ id:y_uuki