Webinar: Untethering Compute from Storage
-
Upload
avere-systems -
Category
Technology
-
view
549 -
download
0
description
Transcript of Webinar: Untethering Compute from Storage
Breaking All The Rules:Untethering Compute from Storage
Bernie Behn
Technical Marketing Engineer
AVERE SYSTEMS, INC5000 McKnight Road, Suite 404
Pittsburgh, PA 15237
(412) 635-7170
averesystems.com
April 23, 2013
Agenda
• Application challenges of data separation anxiety
• Poll: Have you ever… ?• Current methods to overcome these challenges• Poll: How do you … ?• The need for an Edge-Core Architecture• Why the Edge-Core Architecture works• The SPECsfs2008 benchmark test• Avere FXT 3800 Edge filer results• Poll: What do you think … ?• Wrap-up and Q&A
2Proprietary & Confidential
Data Separation Anxiety
• Waiting for I/O is bad! Any process running on a CPU that has to wait for file I/O is stalled
• The further the CPU is from the data it needs, the larger the cost of waste
• Historical approach has been to put the data as close to the CPU as possible, usually on local storage
• NAS file sharing has created the environment where productivity can be increased by enabling parallel computing
3Proprietary & Confidential
The Cost of I/O
L1-cacheL2-cacheRAMDiskNetworkCloud
at 2.4GHz
3 cycles14 cycles
250 cycles41 000 000 cycles
120 000 000 cycles1 200 000 000 cycles
Really? More efficient?
• Sharing data leads to higher efficiency by enabling multiple processing elements to work in parallel
• As access to the shared data takes longer, efficiency gains start to erode
• Spinning up processing farms and/or users at remote sites requires local-like access to data to maintain efficiency
4Proprietary & Confidential
HQ NAS
ComputeFarm HQ
UsersHQ
New RemoteUsers and ComputeWAN
Low Throughput High Latency
Poll #1
NAS protocols like CIFS and NFS were not designed nor intended to effectively handle high-latency/low-bandwidth paths to the fileservers.
• Q: Have you ever tried to access an NFS or CIFS filesystem remotely over a Wide Area Network with high latency and low throughput?
(Please respond within 30 seconds)
5Proprietary & Confidential
• Manual data replication using common tools: rsync/robocopy– Lots of management overhead to keep everything in sync
• Filesystem container mirroring (SnapMirror, SyncIQ, ZFS send/recv)– Often mirroring bulk sets of data, when only subsets of data
are needed
• Just deal with it! It’s not that bad!– What kind of solution is that?
• WAN optimization technologies at the network layer– Why optimize at the network layer when dealing with
Filesystems?
• Host your data in cloud-based storage– Does not solve the latency problem, actually makes it worse
Solutions to this problem?
6Proprietary & Confidential
The Edge-Core NAS Architecture
• Put an end to manual dataset copies and replicate-everywhere tasks
• Use Core filers to centralize data storage management
• Deploy Edge filers to provide efficient access to data
7Proprietary & Confidential
Cloud Storage• Private NAS cloud• Amazon S3 (future)
Primary HQ Datacenter
Cloud Computing• Amazon, Rackspace, SuperNAP
Remote Office
Secondary/Partner/Colo Datacenter
Why Edge-Core solves these problems• Hybrid NAS
– “Cheap and Deep” Core filers for cold and archival data– High performance to “hot dataset” delivered by the Edge filer
• Edge filesystem capabilities with Local Directories– Handle all directory updates at the Edge
• Clustering with linear scalability– Cluster sizes up to 50 nodes, 100+ GByte/sec read throughput
• Data Management with FlashMove™ and FlashMirror™– Handle the moving target of where to stash your at-rest data
8Proprietary & Confidential
Consolidated Management
9Proprietary & Confidential
/
/sales /finance /support
/pipe /staff /fy2012 /cust
NetApp:
EMC/Isilon:
HDS/BlueArc:
Oracle/ZFS:
Data Center (Core Filer:/export)
Remote Site (Core Filer:/export)
/mech
/src
/pipe/cust
/fy2012
/staff
Clients
Avere FXT Edge Filer
WAN
Global Namespace• Simplify namespace mgmt *and* accelerate, scale
perf
• Single mount point for all Core filers & exports
• Easy to add/remove Avere to/from NAS environment
• Create junctions (e.g. /eng) for improved management
/eng
/hw /sw
/mech /src
/ on Avere
Single mount point
FlashMove• Non-disruptively move exports (e.g. /src) between Core
FilersFlashMirror• Mirror write data to two locations for disaster recovery
Global Namespace
/src
FlashMove
/cust’
FlashMirror
X
Logical path unchanged
Poll #2
Shuffling entire datasets, or subsets of data across the globe can be a time and resource intensive activity. It also leads to NAS sprawl and increased costs.
• Q: How do you currently solve your problem of geographically dispersed access to remote filesystems?
(Please respond within 30 seconds)
10Proprietary & Confidential
Performance is King!
• SPECsfs2008 NAS benchmark is designed to exercise all performance and scalability traits of file servers
11Proprietary & Confidential
SFS Aggregate Results for 16 Client(s), Mon Mar 18 16:44:46 2013NFS Protocol Version 3
-------------------------------------------------------------------------------------NFS Target Actual NFS Op NFS Op NFS Mean Std Dev Std Error Pcnt Op NFS NFS Logical Physical Op Response Response of Mean, of Type Mix Mix Success Success Error Time Time 95% Conf Total Pcnt Pcnt Count Count Count Msec/Op Msec/Op +-Msec/Op Time -------------------------------------------------------------------------------------getattr 26.0% 26.0% 124201594 124201594 0 0.54 6.67 0.00 2.3%setattr 4.0% 4.0% 19108469 19108469 0 2.30 27.70 0.00 1.6%lookup 24.0% 24.0% 114650346 114650346 0 0.74 10.41 0.00 3.0%readlink 1.0% 1.0% 4775340 4775340 0 0.56 8.36 0.00 0.1%read 18.0% 18.0% 85985497 105752765 0 18.77 65.52 0.00 58.6%write 10.0% 10.0% 47765007 57779912 0 14.02 61.99 0.00 24.4%create 1.0% 1.0% 4778818 4778818 0 11.39 76.50 0.01 2.0%remove 1.0% 1.0% 4771305 4771305 0 10.78 113.56 0.01 1.9%readdir 1.0% 1.0% 4775784 4775784 0 1.92 23.34 0.00 0.3%fsstat 1.0% 1.0% 4774482 4774482 0 0.54 8.01 0.00 0.1%access 11.0% 11.0% 52556425 52556425 0 0.54 6.29 0.00 1.0%readdirplus 2.0% 2.0% 9557286 9557286 0 14.26 168.35 0.01 4.7%-------------------------------------------------------------------------------------
--------------------------------------------- | SPEC SFS 2008 AGGREGATE RESULTS SUMMARY | ---------------------------------------------SFS NFS THROUGHPUT: 1592334 Ops/Sec AVG. RESPONSE TIME: 5.8 Msec/OpTCP PROTOCOL (IPv4)NFS MIXFILE: [ SFS default ]AGGREGATE REQUESTED LOAD: 1650000 Ops/Sec TOTAL LOGICAL NFS OPERATIONS: 477700353 TEST TIME: 300 Sec TOTAL PHYSICAL NFS OPERATIONS: 507482526PHYSICAL NFS IO THROUGHPUT: 1691608 Ops/secNUMBER OF SFS CLIENTS: 16TOTAL FILE SET SIZE CREATED: 198016272.0 MBTOTAL FILE SET SIZE ACCESSED: 59408280.0 - 60472775.2 MB (100.00% to 101.79% of Base)
Comparing SFS08 MegaOp Solutions*
EMC Isilon$10.7 / IOPS
NetApp$5.1 / IOPS
150ms
Avere$2.3 / IOPS
Avere FXT 3800 Edge filer Cloud Config
13Proprietary & Confidential
Load Gen Client 01
Load Gen Client 02
Load Gen Client 03
Load Gen Client 04
Load Gen Client 05
Load Gen Client 06
Load Gen Client 07
Load Gen Client 08
Load Gen Client 09
Load Gen Client 10
Load Gen Client 11
Load Gen Client 12
Load Gen Client 13
Load Gen Client 14
Load Gen Client 15
Load Gen Client 16
72-port 10 GbE Ethernet Switch
Avere Edge filer 17
Avere Edge filer 18
Avere Edge filer 19
Avere Edge filer 20
Avere Edge filer 21
Avere Edge filer 22
Avere Edge filer 23
Avere Edge filer 24
Avere Edge filer 25
Avere Edge filer 26
Avere Edge filer 27
Avere Edge filer 28
Avere Edge filer 29
Avere Edge filer 30
Avere Edge filer 31
Avere Edge filer 32
Avere Edge filer 01
Avere Edge filer 02
Avere Edge filer 03
Avere Edge filer 04
Avere Edge filer 05
Avere Edge filer 06
Avere Edge filer 07
Avere Edge filer 08
Avere Edge filer 09
Avere Edge filer 10
Avere Edge filer 11
Avere Edge filer 12
Avere Edge filer 13
Avere Edge filer 14
Avere Edge filer 15
Avere Edge filer 16
ZFS Core filer + shelf23 x 4TB SATA
SAS Expansion Shelf23 x 4TB SATA
SAS Expansion Shelf23 x 4TB SATA
10GbE EthernetNetwork path of simulated latency10GbE Core filer with 150ms latencySoftware WAN simulation kernel module
Avere Systems 32 Node SPEC SFS 2008 Configuration
150ms RTT
Poll #3
Benchmarks are a great way to establish apples-to-apples comparisons of available solutions, however, not all workloads are alike. Truth: traditional NAS architectures are not adept to handling large distances between clients and the data they are trying to access.
• Q: What do you think about 1.59 MegaOps of SFS2008 NFS, at 1.24 milliseconds ORT using less than 2 racks of gear with the Core filer 150 milliseconds away?
(Please respond within 30 seconds)14Proprietary & Confidential
Wrap-up / Q&A
• If you are facing any of these challenges, do not despair, there is now an easy and efficient solution!
• Contact [email protected] for more information
• Thank you for attending!
Q&A15Proprietary & Confidential