CERN IT Department CH-1211 Genve 23 Switzerlandt Storage Overview and IT-DM Lessons Learned Luca Canali, IT-DM DM Group Meeting CERN IT Department CH-1211 Genve 23 Switzerlandt Outline Goal: review of storage technology HW layer (HDs, storage array) Interconnect (how to attach storage to the server) Service layer (filesystems) Expose current hot topics in storage Identify challenges Stimulate ideas for management of large data volumes CERN IT Department CH-1211 Genve 23 Switzerlandt Why storage is a very interesting area in the coming years Storage market is very conservative Few vendors share market for large enterprise solutions Enterprise storage has typically a high premium Opportunities Commodity HW/grid-like solutions provide order of magnitude gain in cost/performance New products coming to the market promise many changes: Solid state disks, high capacity disks, high performance and low cost interconnects CERN IT Department CH-1211 Genve 23 Switzerlandt HW layer HD, the basic element Hard disk technology Basic block of storage since 40 years Main intrinsic limitation: latency CERN IT Department CH-1211 Genve 23 Switzerlandt HD specs HDs are limited In particular seek time is unavoidable (7.2k to 15k rpm, ~2-10 ms) IOPS Throughput ~100MB/s, typically limited by interface Capacity range 300GB -2TB Failures: mechanical, electric, magnetic, firmware issues. MTBF: 500k -1.5M hours CERN IT Department CH-1211 Genve 23 Switzerlandt Enterprise disks Performance enterprise disks offer more performance: They spin faster and have better interconnect protocols (e.g. SAS vs SATA) Typically of low capacity Our experience: often not competitive in cost/perf vs. SATA CERN IT Department CH-1211 Genve 23 Switzerlandt HD failure rates Failure rate Our experience: it depends on: vendor, temperature, infant mortality, age. At FAST07 2 papers (one from Google) showed that vendor specs often need to be adjusted in real life. Google data seriously questioned usefulness of SMART probes and correlation of temperature/age/usage with MTBF. Other study showed that consumer and enterprise disks have similar failure pattern and life time. Moreover HD failures in RAID sets have correlations. CERN IT Department CH-1211 Genve 23 Switzerlandt HD wrap-up HD is a old but evergreen technology In particular disk capacities have increased of one order of magnitude in just a few years At the same time prices have gone down (below 0.1 USD per GB for consumer products) 1.5 TB consumer disks, and 450GB enterprise disks are common 2.5 drives are becoming standard to reduce power consumption CERN IT Department CH-1211 Genve 23 Switzerlandt Scaling out the disk The challenge for storage systems Scale out the disk performance to meet demands Throughput IOPS Latency Capacity Sizing storage systems Must focus on critical metric(s) Avoid capacity trap CERN IT Department CH-1211 Genve 23 Switzerlandt RAID and redundancy Storage arrays are the traditional approach implement RAID to protect data. Parity based: RAID5, RAID6 Stripe and mirror: RAID10 Scalability problem of this method For very large configurations MTBF ~ RAID rebuild time (!) Challenge: RAID does not scale CERN IT Department CH-1211 Genve 23 Switzerlandt Beyond RAID Google and Amazon dont use RAID Main idea: Divide data in chunks Write multiple copies of the chunks Google file system: writes chunks in 3 copies Amazon S3: write copies at different destinations, i.e. data center mirroring Additional advantages: Removes the constraint of locally storing redundancy inside one storage arrays Can move, refresh, or relocate data chunks easily CERN IT Department CH-1211 Genve 23 Switzerlandt Our experience Physics DB storage uses ASM Volume manager and cluster file system integrated with Oracle Soon to be also a general-purpose cluster file system (11gR2 beta testing) Oracle files are divided in chunks Chunks are distributed evenly across storage Chunks are written in multiple copies (2 or 3 it depends on file type and configuration) Allows the use of low-cost storage arrays: does not need RAID support CERN IT Department CH-1211 Genve 23 Switzerlandt Scalable and distributed file systems on commodity HW Allow to manage and protect large volumes of data Solutions proven by Google and Amazon, Suns ZFS, Oracles ASM Can provide order of magnitude savings on HW acquisition Additional scale savings by deployment of cloud and virtualization models Challenge: solid and scalable distributed file systems are hard to build CERN IT Department CH-1211 Genve 23 Switzerlandt The interconnect Several technologies available SAN NAS iSCSI Direct attach CERN IT Department CH-1211 Genve 23 Switzerlandt The interconnect Throughput challenge It takes 3 hours to copy/backup 1TB over 1 GBPS network CERN IT Department CH-1211 Genve 23 Switzerlandt IP based connectivity NAS, iSCSI suffer from poor performance of Gbps Ethernet 10 Gbps may/will(?) change the picture At present not widely deployed on servers because of cost Moreover TCP/IP has CPU overhead CERN IT Department CH-1211 Genve 23 Switzerlandt Specialized storage networks SAN is the de facto standard for most enterprise level storage Fast, low overhead on server CPU, easy to configure Our experience (and Tier1s): SAN networks with max 64 ports at low cost Measured: 8 Gbps transfer rate (4+4 dual ported HBAs for redundancy and load balancing) Proof of concept FC backup (LAN free) reached full utilization of tape heads Scalable: proof of concept Oracle supercluster of 410 SATA disks, and 14 dual quadcore servers CERN IT Department CH-1211 Genve 23 Switzerlandt NAS CERNs experience of NAS for databases Netapp filer can use several protocols, the main being NFS Throughput limitation because of TCP/IP Trunking is possible to alleviate the problem, main solution may/will(?) be to move to 10Gbps The filer contains a server with CPU and OS In particular the proprietary WAFL filesystem is capable of creating read-only snapshots Proprietary Data ONTAP OS runs on the filer box Additional features make worse cost/performance CERN IT Department CH-1211 Genve 23 Switzerlandt iSCSI iSCSI is interesting for cost reduction Many concerns on performance though, due to IP interconnect Adoption seems to be only for low-end systems at the moment Our experience: IT-FIO is acquiring some test units, we have been announced that some test HW will be available for IT-DM databases CERN IT Department CH-1211 Genve 23 Switzerlandt The quest for ultimate latency reduction Solid state disks provide unique specs Seek time are at least one order of magnitude better than best HDs A single disk can provide >10k random read IOPS High read throughput CERN IT Department CH-1211 Genve 23 Switzerlandt SSD (flash) problems Flash based SSD still suffer from major problems for enterprise solutions Cost/GB: more than 10 times vs. normal HDs Small capacity compared to HDs They have several issues with write performance Limited number of erase cycles Need to write entire cells (issue for transactional activities) Some workarounds for write performance and cell lifetime improvements are being implemented, different quality from different vendors and grade A field in rapid evolution CERN IT Department CH-1211 Genve 23 Switzerlandt Conclusions Storage technologies are in a very interesting evolution phase On one side old-fashioned storage technologies give more capacity and performance for a lower price every year New technologies are emerging for scaling out very large data sets (see Google, Amazon, Oracles ASM, SUNs ZFS) 10 Gbps Ethernet and SSD have the potential to change storage in the coming years (but are not mature yet) CERN IT Department CH-1211 Genve 23 Switzerlandt Acknowledgments Many thanks to Jacek, Dawid and Maria Eric and Nilo Helge, Tim Bell and Bernd