Ceph Day Beijing - Optimizing Ceph performance by leveraging Intel Optane and 3D NAND TLC SSDs
0 likes112 views
The document discusses optimizing Ceph performance by leveraging Intel Optane and 3D NAND TLC SSDs. It provides an overview of using Optane SSDs as journal/metadata drives and 3D NAND SSDs as data drives in Ceph configurations. It then describes achieving 2.8 million IOPS on an all-flash Ceph cluster with this configuration and analyzes the performance of Optane-based all-flash Ceph arrays. Various tunings are also discussed to improve Ceph performance on Optane drives.
![32
[global]
pid_path = /var/run/ceph
auth_service_required = none
auth_cluster_required = none
auth_client_required = none
mon_data = /var/lib/ceph/ceph.$id
osd_pool_default_pg_num = 2048
osd_pool_default_pgp_num = 2048
osd_objectstore = bluestore
public_network = 172.16.0.0/16
cluster_network = 172.18.0.0/16
enable experimental unrecoverable data
corrupting features = *
bluestore_bluefs = true
bluestore_block_create = false
bluestore_block_db_create = false
bluestore_block_wal_create = false
mon_allow_pool_delete = true
bluestore_block_wal_separate = false
debug objectcacher = 0/0
debug paxos = 0/0
debug journal = 0/0
mutex_perf_counter = True
rbd_op_threads = 4
debug ms = 0/0
debug mds = 0/0
mon_pg_warn_max_per_osd = 10000
debug lockdep = 0/0
debug auth = 0/0
ms_crc_data = False
debug mon = 0/0
debug perfcounter = 0/0
perf = True
debug monc = 0/0
debug throttle = 0/0
debug mds_migrator = 0/0
debug mds_locker = 0/0
[mon]
mon_data = /var/lib/ceph/mon.$id
mon_max_pool_pg_num = 166496
mon_osd_max_split_count = 10000
mon_pg_warn_max_per_osd = 10000
[osd]
osd_data = /var/lib/ceph/mnt/osd-device-$id-data
osd_mkfs_type = xfs
osd_mount_options_xfs = rw,noatime,inode64,logbsize=256k
bluestore_extent_map_shard_min_size = 50
bluefs_buffered_io = true
mon_osd_full_ratio = 0.97
mon_osd_nearfull_ratio = 0.95
bluestore_rocksdb_options =
compression=kNoCompression,max_write_buffer_number=32,min_write_buffer_number_to_merge
=2,recycle_log_file_num=32,compaction_style=kCompactionStyleLevel,write_buffer_size=6710886
4,target_file_size_base=67108864,max_background_compactions=31,level0_file_num_compaction
_trigger=8,level0_slowdown_writes_trigger=32,level0_stop_writes_trigger=64,num_levels=7,max_b
ytes_for_level_base=536870912,max_bytes_for_level_multiplier=8,compaction_threads=32,flusher
_threads=8
bluestore_min_alloc_size = 65536
osd_op_num_threads_per_shard = 2
osd_op_num_shards = 8
bluestore_extent_map_shard_max_size = 200
bluestore_extent_map_shard_target_size = 100
bluestore_csum_type = none
bluestore_max_bytes = 1073741824
bluestore_wal_max_bytes = 2147483648
bluestore_max_ops = 8192
bluestore_wal_max_ops = 8192
Ceph All Flash Tunings
debug rgw = 0/0
debug finisher = 0/0
debug osd = 0/0
debug mds_balancer = 0/0
rocksdb_collect_extended_stats = True
debug hadoop = 0/0
debug client = 0/0
debug zs = 0/0
debug mds_log = 0/0
debug context = 0/0
rocksdb_perf = True
debug bluestore = 0/0
debug bluefs = 0/0
debug objclass = 0/0
debug objecter = 0/0
debug log = 0
ms_crc_header = False
debug filer = 0/0
debug rocksdb = 0/0
rocksdb_collect_memory_stats = True
debug mds_log_expire = 0/0
debug crush = 0/0
debug optracker = 0/0
osd_pool_default_size = 2
debug tp = 0/0
cephx require signatures = False
cephx sign messages = False
debug rados = 0/0
debug journaler = 0/0
debug heartbeatmap = 0/0
debug buffer = 0/0
debug asok = 0/0
debug rbd = 0/0
rocksdb_collect_compaction_stats = False
debug filestore = 0/0
debug timer = 0/0
rbd_cache = False
throttler_perf_counter = False](https://image.slidesharecdn.com/3-170613192359/85/Ceph-Day-Beijing-Optimizing-Ceph-performance-by-leveraging-Intel-Optane-and-3D-NAND-TLC-SSDs-32-320.jpg)
Ceph Day Beijing - Optimizing Ceph performance by leveraging Intel Optane and 3D NAND TLC SSDs
- 1. OptimizingCephPerformancebyLeveraging Intel®Optane™and3DNANDTLCSSDs Yuan Zhou, Software Engineer, [email protected] Jack Zhang, Storage Solution Architect, [email protected] Jun, 2017
- 2. 2 Agenda • Ceph* with Intel® Non-Volatile Memory Technologies • Ceph* @ PRC • 2.8M IOPS Ceph* cluster with Intel® Optane™ SSDs + Intel® 3D TLC SSDs • Ceph* Performance analysis on Intel® Optane™ SSDs based all-flash array • Summary
- 3. 3
- 4. 4NVM Solutions Group Intel Confidential Intel®3DNANDSSDsandOPTANESSDTransformStorage Expand the reach of Intel® SSDs. Deliver disruptive value to the data center. 11 Refer to appendix for footnotes Capacity for Less Performance for Less Optimized STORAGE
- 5. 5 InnovationforCloudSTORAGE: Intel® Optane™ + Intel® 3D NAND SSDs New Storage Infrastructure: enable high performance and cost effective storage: Journal/Log/Cache Data Openstack/Ceph: ‒ Intel Optane™ as Journal/Metadata/WAL (Best write performance, Lowest latency and Best QoS) ‒ Intel 3D NAND TLC SSD as data store (cost effective storage) ‒ Best IOPS/$, IOPS/TB and TB/Rack Ceph Node (Yesterday) P3520 2TB P3520 2TB P3520 2TB P3520 2TB P3700 U.2 800GB Ceph Node (Today) Intel® Optane™ P4800X (375GB) 3D NAND P4500 4TB 3D XPoint™ P4500 4TB P4500 4TB P4500 4TB P4500 4TB P4500 4TB Transition to + P4500 4TB P4500 4TB
- 6. Suggested Configurations for Ceph* Storage Node Standard/good (baseline): Use cases/Applications: that need high capacity storage with high throughput performance NVMe*/PCIe* SSD for Journal + Caching, HDDs as OSD data drive Better IOPS Use cases/Applications: that need higher performance especially for throughput, IOPS and SLAs with medium storage capacity requirements NVMe/PCIe SSD as Journal, High capacity SATA SSD for data drive Best Performance Use cases/Applications: that need highest performance (throughput and IOPS) and low latency/QoS (Quality of Service). All NVMe/PCIe SSDs More information at Ceph.com (new RAs update soon!) http://tracker.ceph.com/projects/ceph/wiki/Tuning_for_All_Flash_ Deployments *Other names and brands may be claimed as the property of others. 6 Ceph* storage node --Good CPU Intel(R) Xeon(R) CPU E5-2650v4 Memory 64 GB NIC 10GbE Disks 1x 1.6TB P3700 + 12 x 4TB HDDs (1:12 ratio) P3700 as Journal and caching Caching software Intel(R) CAS 3.0, option: Intel(R) RSTe/MD4.3 Ceph* Storage node --Better CPU Intel(R) Xeon(R) CPU E5-2690v4 Memory 128 GB NIC Duel 10GbE Disks 1x Intel(R) DC P3700(800G) + 4x Intel(R) DC S3510 1.6TB Or 1xIntel P4800X (375GB) + 8x Intel® DC S3520 1.6TB Ceph* Storage node --Best CPU Intel(R) Xeon(R) CPU E5-2699v4 Memory >= 128 GB NIC 2x 40GbE, 4x dual 10GbE Disks 1xIntel P4800X (375GB) + 6x Intel® DC P4500 4TB
- 7. 7 Who is using Ceph? Searchable Examples: http://www.slideshare.net/inktank_ceph Telcom CSP/IPDC OEM/ODM Enterprise, FSI, Healthcare, Retailers
- 8. 8 Ceph* @ PRC Ceph* is very hot in PRC Redevelopment based on the upstream code More and more companies move to OpenSource storage solutions Intel/Redhat held Three Ceph days at Beijing and Shanghai 1000+ attendees from 500+ companies Self media-reports and re-clippings Next one: Ceph Day Beijing on June, 10th More and more PRC code contributors ZTE, XSKY, H3C, LETV, UnitedStack, AliYun, Ebay, EasyStack
- 9. 9 Ceph* on all-flash array Storage providers are struggling to achieve the required high performance There is a growing trend for cloud providers to adopt SSD – CSP who wants to build EBS alike service for their OpenStack* based public/private cloud Strong demands to run enterprise applications OLTP workloads running on Ceph, tail latency is critical high performance multi-purpose Ceph cluster is a key advantage Performance is still an important factor SSD price continue to decrease *Other names and brands may be claimed as the property of others.
- 10. 10
- 11. 11 Ceph* performance trend with SSD – 4K Random Write 38x performance improvement in Ceph all-flash array! 1.98x 3.7x 1.66x 1.23x 1.19x
- 12. CEPH2 12 Ceph* All Flash Optane configuration 8x Client Node • Intel® Xeon™ processor E5-2699 v4 @ 2.3GHz, 64GB mem • 1x X710 40Gb NIC 8x Storage Node • Intel Xeon processor E5-2699 v4 @ 2.3 GHz • 256GB Memory • 1x 400G SSD for OS • 1x Intel® DC P4800 375G SSD as WAL and DB • 8x 2.0TB Intel® SSD DC P4500 as data drive • 2 OSD instances one each P4500 SSD • Ceph 12.0.0 with Ubuntu 16.10 2x40Gb NIC Test Environment CEPH1 MON OSD1 OSD16… FIO FIO CLIENT 1 1x40Gb NIC FIO FIO CLIENT 2 FIO FIO CLIENT 3 FIO FIO ….. FIO FIO CLIENT8 *Other names and brands may be claimed as the property of others. CEPH3 … CEPH8 Workloads • Fio with librbd • 20x 30 GB volumes each client • 4 test cases: 4K random read & write; 64K Sequential read & write
- 13. 13 Excellent performance on Optane cluster – random read & write hit CPU bottleneck Ceph* Optane Performance overview Throughput Latency (avg.) 99.99% latency 4K Random Read 2876K IOPS 0.9 ms 2.25 4K Random Write 610K IOPS 4.0 ms 25.435 64K Sequential Read 27.5 GB/s 7.6 ms 13.744 64K Sequential Write 13.2 GB/s 11.9 ms 215
- 14. 14 • Good Node Scalability, poor disk scalability for 4K block workloads (CPU throttled!) • NUMA & HT helps a lot on the Performance • Fine tune the # of OSD per node and Drive per Node. Ceph* Optane Performance – Tunings
- 15. 15 Ceph* Optane Performance – Performance improvement • The breakthrough high performance of Optane eliminated the WAL & rocksdb bottleneck • 1 P4800X or P3700 covers up to 8x P4500 data drivers as both WAL and rocksdb
- 16. 16 Ceph* Optane Performance – Rocksdb improvement • Eliminate rocksdb write stall with Optane drive • Added event listener in rocksdb to provide real-time compaction&&flush information • Write stalls when flush or compaction can’t keep up with the incoming write rate • rocksdb_compaction matches submit_transaction latency increase points
- 17. 17 Ceph* Optane Performance – latency improvement • Significant tail latency improvement with Optane • 20x latency reduction for 99.99% latency 5.7 ms 14.7 ms 18.3 ms 317.8 ms
- 18. 18
- 19. 19 Ceph* Optane Performance Analysis - CPU utilizations • Random Read & Write performance are throttled by CPU • Unbalanced CPU utilization caused by HT efficiency for random workloads • Limiting # of drive scalability for small block random workloads • Need to optimize CPU utilization
- 20. Ceph* Optane Performance Analysis - CPU profiling for 4K RR • Perf record for 30 second • Ceph-osd: 34.1%, tp_osd_tp: 65.6% • Heavy network messenger overhead AsyncMsg 20
- 21. 21 Ceph* Optane Performance Analysis - CPU profiling for 4K RR
- 22. Ceph* Optane Performance Analysis - CPU profiling for 4K RR • The top three consumers of tp_osd_tp thread are • KernelDevice::read, • OSD::ShardedOpWQ::_ process • PrimaryLogPG::do_op. • Perf record 30s. 22
- 23. Ceph* Optane Performance Analysis - CPU profiling for 4K RW Wal remove thread(~1.85%) RocksDB(about ~6%-~7% ) AsyncMsg(~28.9%) 23
- 24. Ceph* Optane Performance Analysis - CPU profiling for 4K RW BlueStore(~32% - ~36%) 24
- 25. Ceph* Optane Performance Analysis - WAL tunings and optimizations • Key takeaways: • rocksdb memTable overhead lead to 50000 IOPS difference. (tuning2 vs. tuning3) • If we don’t clean bluestore WAL data in rocksdb, rocksdb overhead increase dramatically with the # of metadata increase: • Use an external WAL device to store WAL data, and just write WAL metadata into rocksdb! Tunings 4K RW IOPS comments Default: Baseline(NVMe as DB&& WAL drive) 340000 Separated DB&&WAL device Tuning1: DB on NVMe && WAL on Ramdisk 360000 Move WAL to Ramdisk Tuning2: Tuning1+disable rocksdb WAL 360000 RocksDB tuning Tuning3: Tuning2+omit WAL in deferred write mode 410000 Don’t write WAL in deferred write mode Tuning4: Tuning1+write WAL but don’t remove WAL from rocksdb 240000 Write WAL before write metadata into rocksdb, but will not clean WAL after write data to data device in deferred write mode Tuning5: Tuning1+external WAL 380000 Write WAL to an external WAL device, and write its metadata to rocksdb 25
- 26. 26 Ceph* Optane Performance Analysis - Direct I/O vs Buffered I/O • Key takeaways: • On P3700 setup, with bulefs_buffered_io, the performance improved 9%, 99.99% latency improved ~8x • However on Optane setup, we find bluefs_direct_io could improve the 10% performance, 99.99% latency also got improved ~1x 0.00x 0.02x 0.06x 0.25x 1.00x 0.00x 0.20x 0.40x 0.60x 0.80x 1.00x 1.20x 1.40x bluefs_bufferd_io bluefs_direct_io 99.99%_lat(ms) IOPS 4k rand write performance: buffered vs direct P3700 iops Optane iops P3700 99.99%_lat Optane 99.99%_lat
- 27. 27 Ceph* Optane Performance Analysis - Metadata plane I/O pattern DB WAL • Key takeaways: • Sync writes! • Large sequential write w/ offset-overlapping on WAL/DB device • Small disk reads on DB device due to metadata lookup missing on writes • BlueFS will save tail part of previous unaligned writes and merge with current writes • Medium-sized requests due to db transaction batch in BlueStore
- 28. 28
- 29. 29 Summary & Next Summary • Ceph* is awesome! • Strong demands for all-flash array Ceph* solutions • Optane based all-flash array Ceph* cluster is capable of delivering over 2.8M IOPS with very low latency! • Let’s work together to make Ceph* more efficient with all-flash array! Next • Client side cache on Optane with SQL workloads! *Other names and brands may be claimed as the property of others.
- 30. 30 Acknowledgements This is a joint team work. Thanks for the contributions of Haodong, Tang, Jianpeng Ma.
- 31. 31 Backup
- 32. 32 [global] pid_path = /var/run/ceph auth_service_required = none auth_cluster_required = none auth_client_required = none mon_data = /var/lib/ceph/ceph.$id osd_pool_default_pg_num = 2048 osd_pool_default_pgp_num = 2048 osd_objectstore = bluestore public_network = 172.16.0.0/16 cluster_network = 172.18.0.0/16 enable experimental unrecoverable data corrupting features = * bluestore_bluefs = true bluestore_block_create = false bluestore_block_db_create = false bluestore_block_wal_create = false mon_allow_pool_delete = true bluestore_block_wal_separate = false debug objectcacher = 0/0 debug paxos = 0/0 debug journal = 0/0 mutex_perf_counter = True rbd_op_threads = 4 debug ms = 0/0 debug mds = 0/0 mon_pg_warn_max_per_osd = 10000 debug lockdep = 0/0 debug auth = 0/0 ms_crc_data = False debug mon = 0/0 debug perfcounter = 0/0 perf = True debug monc = 0/0 debug throttle = 0/0 debug mds_migrator = 0/0 debug mds_locker = 0/0 [mon] mon_data = /var/lib/ceph/mon.$id mon_max_pool_pg_num = 166496 mon_osd_max_split_count = 10000 mon_pg_warn_max_per_osd = 10000 [osd] osd_data = /var/lib/ceph/mnt/osd-device-$id-data osd_mkfs_type = xfs osd_mount_options_xfs = rw,noatime,inode64,logbsize=256k bluestore_extent_map_shard_min_size = 50 bluefs_buffered_io = true mon_osd_full_ratio = 0.97 mon_osd_nearfull_ratio = 0.95 bluestore_rocksdb_options = compression=kNoCompression,max_write_buffer_number=32,min_write_buffer_number_to_merge =2,recycle_log_file_num=32,compaction_style=kCompactionStyleLevel,write_buffer_size=6710886 4,target_file_size_base=67108864,max_background_compactions=31,level0_file_num_compaction _trigger=8,level0_slowdown_writes_trigger=32,level0_stop_writes_trigger=64,num_levels=7,max_b ytes_for_level_base=536870912,max_bytes_for_level_multiplier=8,compaction_threads=32,flusher _threads=8 bluestore_min_alloc_size = 65536 osd_op_num_threads_per_shard = 2 osd_op_num_shards = 8 bluestore_extent_map_shard_max_size = 200 bluestore_extent_map_shard_target_size = 100 bluestore_csum_type = none bluestore_max_bytes = 1073741824 bluestore_wal_max_bytes = 2147483648 bluestore_max_ops = 8192 bluestore_wal_max_ops = 8192 Ceph All Flash Tunings debug rgw = 0/0 debug finisher = 0/0 debug osd = 0/0 debug mds_balancer = 0/0 rocksdb_collect_extended_stats = True debug hadoop = 0/0 debug client = 0/0 debug zs = 0/0 debug mds_log = 0/0 debug context = 0/0 rocksdb_perf = True debug bluestore = 0/0 debug bluefs = 0/0 debug objclass = 0/0 debug objecter = 0/0 debug log = 0 ms_crc_header = False debug filer = 0/0 debug rocksdb = 0/0 rocksdb_collect_memory_stats = True debug mds_log_expire = 0/0 debug crush = 0/0 debug optracker = 0/0 osd_pool_default_size = 2 debug tp = 0/0 cephx require signatures = False cephx sign messages = False debug rados = 0/0 debug journaler = 0/0 debug heartbeatmap = 0/0 debug buffer = 0/0 debug asok = 0/0 debug rbd = 0/0 rocksdb_collect_compaction_stats = False debug filestore = 0/0 debug timer = 0/0 rbd_cache = False throttler_perf_counter = False
- 33. Legalnotices Copyright © 2016 Intel Corporation. All rights reserved. Intel, the Intel logo, Xeon, Intel Inside, and 3D XPoint are trademarks of Intel Corporation in the U.S. and/or other countries. *Other names and brands may be claimed as the property of others. FTC Optimization Notice Intel's compilers may or may not optimize to the same degree for non-Intel microprocessors for optimizations that are not unique to Intel microprocessors. These optimizations include SSE2, SSE3, and SSSE3 instruction sets and other optimizations. Intel does not guarantee the availability, functionality, or effectiveness of any optimization on microprocessors not manufactured by Intel. Microprocessor-dependent optimizations in this product are intended for use with Intel microprocessors. Certain optimizations not specific to Intel microarchitecture are reserved for Intel microprocessors. Please refer to the applicable product User and Reference Guides for more information regarding the specific instruction sets covered by this notice. Notice revision #20110804 The cost reduction scenarios described in this document are intended to enable you to get a better understanding of how the purchase of a given Intel product, combined with a number of situation-specific variables, might affect your future cost and savings. Nothing in this document should be interpreted as either a promise of or contract for a given level of costs. Tests document performance of components on a particular test, in specific systems. Differences in hardware, software, or configuration will affect actual performance. Consult other sources of information to evaluate performance as you consider your purchase. For more complete information about performance and benchmark results, visit http://www.intel.com/performance. 33