Understanding the Impact of Cache Locations on Storage ... - Usenix
Recommend Documents
cache access involves the costly I/O virtualization layers, which cause ... intensive VMs, the hypervisor-side caching consumes .... To fetch the data, QEMU issues I/O requests which .... side caches; for 1MB large I/O requests, the maximum.
time, calculated as the amount of time it took to return a ... better than the Python matrix queries (8x), PMGD per- ... In the figure, python matrix, memsql, and.
Jun 20, 2014 - the SSD interface has been modified for caching pur- poses [26], but ... plausible primary storage system with Nitro acceleration. ... I/O rates for.
Storage over IP: When Does Hardware Support help? Prasenjit Sarkar ... build an IP storage area network that differ in their level of hardware support: software, TOE (TCP Offload Engine) and HBA .... tems that provide remote file access over IP netwo
of a storage system, within the hardware and firmware of disk enclosures ... providers for storage and processing (e.g., Web mail,. DropBox, Amazon S3). Most of ...
Aug 16, 2017 - tial attack on Krebs exceeded 600 Gbps in volume [46]â among the largest ... peak attack. Figure 1: Mir
Aug 16, 2017 - Manos Antonakakis, Georgia Institute of Technology; Tim April, Akamai; Michael Bailey,. University of Ill
Jul 22, 2015 - and laptops devices. In particular, it is much ... Several device manufactures, such as Apple and Samsung, have recently introduced .... A passcode can be either (1) a simple 4-digit PIN1 or (2) a longer one, with up to 37 ...
Jul 22, 2015 - non-Touch ID group. Note, that those iPhone 5S, 6 and, 6 Plus ...... fingerprint authentication versus pins: A usability study. (cmu-cylab-14-012).
F or instance, some of these attacks are dedicated to embedded devices, some others .... This means that when timing a sequence of memory accesses, the number .... request to the targeted server ( Alice ) using Alice's public key (F igure 4) .
Don't Thrash: How to Cache your Hash on Flash. Michael A. ... do not scale outside of main memory. ... bles do not contain the queried element, meaning that the ..... html. [12] G. J. Holzmann. Design and validation of computer pro- tocols.
Nov 13, 2015 - As a rule of thumb, industry practitioners use 70% of the maximum IOPS as the 'knee of the curve'. ..... T. J., and Wynn, H. P. Design and analy-.
Jul 22, 2015 - free apps and 32% of paid apps provide in-app access to a privacy policy [8]. ..... Top-100 domain-specific stop words for English and Chinese.
Nov 13, 2015 - as these, storage workloads vary reasonably and ... tions, storage providers often over-provision stor- ... in cloud provider environment. Usually ...
Jul 22, 2015 - ours is given in section 7 (see Table 10 for a quick summary), and here we note .... To check the existen
Dec 3, 2007 - 2007 Intel Corporation. Impact of Cache Coherence Protocols on the Processing of Network Traffic. Amit Kumar and Ram Huggahalli.
for compressing the instruction stream for programs ... An I-cache miss will stall the processor, while the ..... tures, ACM Distinguished Dissertation, MIT. Press ...
$1;. } The Impact of Instruction Compression on. I-cache Performance. I-Cheng K. Chen. Peter L. Bird. Trevor Mudge. EECS Department. University of Michigan.
6 Jun 2005 - replacement algorithm is the Least Recently Used (LRU) re- placement policy [9]. The effectiveness of LRU comes from the simple yet powerful ...
297-676537; Email: [email protected] ... counterpart. Given the link between mimicry and the desire to create affiliation, we hypothesized that guilt ...
Intermediate Gateway Service to Aggregate and Cache the I/O Operations into.
Distributed Storage Repositories. Mehmet Balman, Ismail Akturk, and Tevfik ...
cache miss directed to the faulty disk, and thus to reduce the I/O requests to the ... In this section we briefly overview some background ma- terials and related ...
Fac. of Electrical Engineering, Mathematics and Computer Science,. Univ. of Twente ...... be kept in mind that a forensic investigation carried out by the police is ...
coloring substances in juices and syrups. ... The sugar beet was stored at different ranges of temperature of 2, 6, 10, ... Therefore, the quantity of inverted sugar.
Understanding the Impact of Cache Locations on Storage ... - Usenix
Overview. ⢠Cloud storage stacks. ⢠Impact of cache locations on system performance and energy efficiency. ⢠Potential optimizations. 11 ...
Understanding the Impact of Cache Locations on Storage Performance and Energy Consumption of Virtualization Systems Tao Lu, Ping Huang, and Xubin He Virginia Commonwealth University Ming Zhang EMC Corporation
Caching: one key frees two birds Server load
I/O latency
2
Caching: one key frees two birds Server load Poorly cached
I/O latency
3
Caching: one key frees two birds Server load
Well cached I/O latency
4
Cache devices DRAM PCIe SSD
SATA SSD
5
Cloud caching: a puzzle VM
VM
Hypervisor
VM
VM
Hypervisor
VM
VM
Hypervisor
Where should the cache be Network deployed? Storage
6
Cloud caching: traditional wisdom
Ioan Stefanovici et. al., Software-Defined Caching: Managing Caches in Multi-Tenant Data Centers, SoCC’15 7
Cloud caching: traditional wisdom (cont’d) • D. Arteaga et. al., Cloudcache: On-demand Flash Cache Management for Cloud Computing , FAST’16 • Hwang, J. et.al., Mortar: Filling the Gaps in Data Center Memory, VEE’14 • Byan, S. et.al., Mercury : Host-side Flash Caching for the Data Center, MSST’12 8
Cloud caching: traditional wisdom VM
VM
VM
Hypervisor
VM
VM
Hypervisor
VM
Hypervisor
Caches are deployed in hypervisors Storage Network
9
Is hypervisor the best location to deploy caches?
10
Overview • Cloud storage stacks • Impact of cache locations on system performance and energy efficiency • Potential optimizations
11
Overview • Cloud storage stacks • Impact of cache locations on system performance and energy efficiency • Potential optimizations
Run fio micro-benchmark on a physical machine ~2us • Run fio once to warm up the page cache • Run fio for a second time to measure the page cache performance Run fio on a physical machine ~10us • Build a RAMDisk • Run fio on the RAMDisk with O_DIRECT enabled to bypass the page cache
Run fio on a physical machine 100+ us • Run fio on a SSD/HDD with O_DIRECT enabled to bypass the page cache
Overview • Cloud storage stacks • Impact of cache locations on system performance and energy efficiency • Potential optimizations
16
Testbed VM
Host &. Hypervisor
Storage
17
Testbed VM Host &. Hypervisor Storage
● Physical machine and Host OS ○ AMD Phenom II X4 B95 Quad-core 3.0 GHz processor with AMD-V virtualization support ○ 8GB DRAM ○ Host OS is a 64-bit Ubuntu 15.04 with Linux kernel version 3.19.0- 30-generic. ○ QEMU emulator version 2.4.1 and KVM are used as the hypervisor. 18
Testbed VM Host &. Hypervisor
● VM & Guest OS ○ 2 VCPUs and 2GB memory ○ Ubuntu 15.04 64-bit Server Cloud Image as the guest OS
Storage
19
Testbed VM Host &. Hypervisor Storage
● Disks ○ One 500GB Samsung 850 Pro SATA SSD ○ One 240GB OCZ RevoDrive 3 PCIe SSD
20
Testbed VM Host &. Hypervisor
● Benchmark ○ fio
Storage
21
DRAM caches: VM-side vs. Hypervisor-side vs. Native Benchmark
Benchmark Benchmark
VM
VM
Host &. Hypervisor
Host &. Hypervisor
Host &. Hypervisor
Storage
Storage
Storage
22
Performance of DRAM caches Benchmark
VM Hypervisor Storage
450000
Benchmark
250000
Benchmark
Hypervisor Storage
350000
Native VM-side Hypervisor-side
11%
300000
IOPS
VM Hypervisor Storage
400000
97%
200000 150000 100000 50000 0 4
8
16
32
64
128
256
512
1024
I/O size (KB) 23
Power consumption of DRAM caches Benchmark
Benchmark
VM Hypervisor Storage Benchmark
Hypervisor Storage
30
Power consumption (watts)
VM Hypervisor Storage
Native
VM-side
Hypervisor-side
25
1.7x
20 15
10
0.9x
9x
5 0 1
2
3
4
5
6
7
8
9
10
IOPS (K) 24
For DRAM caches, VM-side caching has better performance and energy efficiency than Hypervisor-side caching. 25
DRAM cache comparison: VM-side vs. Hypervisor-side vs. Native Benchmark
Benchmark Benchmark
VM
Host &. Hypervisor Storage
VM KVM DRAM
Host &. Hypervisor Storage
virtio DRAM
Host &. Hypervisor
DRAM
Storage
26
SATA SSD caches: VM-side vs. Hypervisor-side vs. Native Benchmark
Benchmark
Benchmark
VM
VM
Host &. Hypervisor
Host &. Hypervisor
SATA SSD
SATA SSD
Host &. Hypervisor SATA SSD 27
Performance of SATA SSD caches
28
Power consumption of SATA SSD caches
29
For SATA SSD based caches, VM-side caching and Hypervisor-side caching perform similarly.
30
SATA SSD cache comparison: VM-side vs. Hypervisor-side vs. Native Benchmark
Benchmark
Benchmark
VM
VM
Host &. virtio Hypervisor
virtio Host &. Hypervisor
SATA SSD
SATA SSD
Host &. Hypervisor SATA SSD
31
PCIe SSD caches: PCI passthrough vs. virtio Benchmark
Benchmark
VM
VM
Host &. PCI-pass Hypervisor PCIe SSD
Host &. virtio Hypervisor PCIe SSD 32
Performance of PCIe SSD caches (cont’d) Benchmark
VM Host &. PCI-pass Hypervisor PCIe SSD Benchmark
VM Host &. virtio Hypervisor PCIe SSD
33
Power consump. of PCIe SSD caches (cont’d) Benchmark
VM Host &. PCI-pass Hypervisor PCIe SSD Benchmark
VM Host &. virtio Hypervisor PCIe SSD
34
PCI passthrough enables the PCIe SSD allocated to a VM achieving near native performance.
35
Is hypervisor the best location to deploy caches? SATA SSD Maybe YES
PCIe SSD NO
DRAM Absolutely NO
36
What caps the performance of hypervisor-side caching? Virtual I/O path
![Understanding the Inconsistencies between Text ... - Usenix [PDF]](https://m.moam.info/img/260x300/understanding-the-inconsistencies-between-text-use_648a6f7e098a9e3d5e8b458f.jpg)
