Embedded System Lab DCStore Eliminating Noisy Neighbor Containers
Embedded System Lab. DC-Store: Eliminating Noisy Neighbor Containers using Deterministic I/O Performance and Resource Isolation Miryeong Kwon et al. , USENIX FAST, 2020. 06. 08 Presentation by Sopanhapich CHUM sopanhapich. chum@gmail. com
Embedded System Lab. Content 1. Introduction 2. Multi-Container Scenario 3. Divided SSD: Hardware Level Design 4. I/O Tacker: System Level Support 5. Evaluation 6. Conclusions 2
1. Introduction Kernel-space User-space Ø Motivation: When sharing storage in multi-container environment, a write-intensive container can be a noisy neighbor for other read-intensive containers Hardware (Storage) § Embedded System Lab. Read-intensive container Write-intensive container Reads Writes Isolation I/O Tacker (System-level) Divided SSD (HW-level) 3
2. Multi-Container Scenario § Embedded System Lab. Facebook: Representative Example of Multi-Container “Interspersing bursty write into sustained reads” [Flash Memory Summit’ 18] Goal of storage sharing : service reads without impact of writes 4
2. Multi-Container Scenario Facebook: Representative Example of Multi-Container read-intensive write-intensive Room #2 Room #1 § Embedded System Lab. 5
2. Multi-Container Scenario Embedded System Lab. Soundproof test of Logically Separated Volume § § § Let’s check how much the logically separated room can soundproof Co-run a read-intensive container with three of write-intensive container (fio) A B C D Could A listen to music in a quiet environment? 6
2. Multi-Container Scenario § Embedded System Lab. Container A’s User experience 7
3. Divided SSD: Hardware Level Design § Embedded System Lab. HW-level Solution: Divided SSD. Ø Idea: Let’s physically separate all internal HW resources , sitting on the critical path of I/O services Ø Our Divided SSD can provide physically separated storage volume which is called NVM Set #1 NVM Set #2 NVM Set #3 NVM Set #4 Divided SSD 8
3. Divided SSD: Hardware Level Design § Embedded System Lab. NVM set design 9
4. I/O Tacker: System Level Support § Embedded System Lab. Background write I/O 10
4. I/O Tacker: System Level Support § Embedded System Lab. Limitation of Conventional Noise Elimination Ø Conventional consideration of noisy sound : Assume that only the foreground I/O request generated by running applications are the noisy sound Ø Observation: Background I/O (caused by OS management of metadata) are practically load noisy sounds Redefined noisy sound = Foreground writes + Background writes 11
4. I/O Tacker: System Level Support § Embedded System Lab. Page Frame Reclaiming: Representative Example Ø The representative example of background noisy sound is page frame reclaiming 12
4. I/O Tacker: System Level Support § Embedded System Lab. How Page Frame Reclaiming Works in Linux Ø The representative example of background noisy sound is page frame reclaiming container A’s page frame A D C B B’s page frame C’s page frame D’s page frame Swap area NVM set A NVM set B NVM set C NVM set D 13
4. I/O Tacker: System Level Support § Embedded System Lab. Pinning Metadata Destination container A’s page frame A D C B B’s page frame C’s page frame D’s page frame Swap area NVM set A NVM set B NVM set C NVM set D 14
4. I/O Tacker: System Level Support Embedded System Lab. Ø I/O Tacker mainly provides three system-level support container A’s page frame 3. Pinned information aware page frame reclaiming 2. Container creation with metadata area pinning information A Metadata = NVM Set A 1. Sharable/Pinnable Swap Area Property Pinnable swap area A NVM Set B 15
5. Evaluation § Embedded System Lab. Experiment Setup Ø The hardware configuration of Baseline SSD and Divided SSD prototype are exactly the same as each other CPU Intel Xeon CPU E 5 -2690 v 3 @2. 60 GHz, 48 Processors, 12 Cores, 30, 720 KB cached size Configuration Capacity 2 TB #channel 8 #core 48 DRAM 256 GB #die 64 NAND 64 Layered TLC 16
5. Evaluation § Embedded System Lab. Experiment Setup Ø Let’s redefine the type of applications based on the I/O access pattern than a container generated § Frequently generates I/O request Ø grep Ø wordcout Ø minmax Memory-Hungry Compute-intensive Data-intensive § Rarely generates I/O request Ø Sort § Noisy generates I/O request Ø LEMP stack (Full stack of DB) 17
5. Evaluation Experiment Setup Memory-Hungry Baseline Data-intensive or Compute-intensive Swap area Data-intensive or Compute-intensive Memory-Hungry DC-Store § Embedded System Lab. Swap area 18
5. Evaluation Execution Time comparison Memory-Hungry Baseline Data-intensive or Compute-intensive Swap area Data-intensive or Compute-intensive Memory-Hungry DC-Store § Embedded System Lab. Swap area 19
5. Evaluation Time Series Analysis Memory-Hungry Baseline Data-intensive or Compute-intensive Swap area Data-intensive or Compute-intensive Memory-Hungry DC-Store § Embedded System Lab. Swap area 20
5. Evaluation I/O Tacker Performance Memory-Hungry Default Data-intensive or Compute-intensive Swap area Round Robin Data-intensive or Compute-intensive Swap area Data-intensive or Compute-intensive Memory-Hungry Swap area Memory-Hungry Pinning § Embedded System Lab. Swap area 21
5. Evaluation I/O Tacker Performance Memory-Hungry Default Data-intensive or Compute-intensive Swap area Round Robin Data-intensive or Compute-intensive Swap area Data-intensive or Compute-intensive Memory-Hungry Swap area Memory-Hungry Private § Embedded System Lab. Swap area 22
5. Evaluation I/O Tacker Performance Memory-Hungry Default Data-intensive or Compute-intensive Swap area Round Robin Data-intensive or Compute-intensive Swap area Data-intensive or Compute-intensive Memory-Hungry Swap area Memory-Hungry Private § Embedded System Lab. Swap area 23
6. Conclusion Embedded System Lab. § Conclusion • DC-Store successfully eliminated the noisy I/O request. • Prototyped both hardware (Divided SSD) and software (I/O Tacker), evaluate them in a real system which shortens the data-intensive applications by 31% on average compare to those on a baseline system. 24
Embedded System Lab. DC-Store: Eliminating Noisy Neighbor Containers using Deterministic I/O Performance and Resource Isolation Miryeong Kwon et al. , USENIX FAST, 2020 Thank You! 2020. 06. 08 Presentation by sopanhapich CHUM sopanhapich. chum@gmail. com
- Slides: 25