Performance Characteristics of Large Analytical Query Processing on Relational Database with SSDs

Transcription

1 THE INSTITUTE OF ELECTRONICS, INFORMATION AND COMMUNICATION ENGINEERS TECHNICAL REPORT OF IEICE. SSD,,, SSD HDD 1 1 HDD SSD HDD HDD SSD SSD, SSD, OLAP, Performance Characteristics of Large Analytical Query Processing on Relational Database with SSDs Keisuke SUZUKI, Yuto HAYAMIZU, Daisaku YOKOYAMA, Miyuki NAKANO, and Masaru KITSUREGAWA, Kitsuregawa Lab., Center for Information Fusion, Institute of Industrial Science (IIS), Meguro-ku Komaba 4-6-1, Tokyo, Japan, Hitotsubashi, Chiyoda-ku, Tokyo, Japan, Abstract The SSDs are expected new storage devices whose transfer rate are thousands faster than that of HDDs. By employing SSDs on behalf of HDDs, performance of data intensive applications is likely to be improved drastically since I/O cost of HDDs is dominant in execution time of data intensive applications. Then, different behavior of data intensive application may be observed in case of SSDs, since SSD s I/O cost is relatively small in total execution time. In this report, we analyze the difference in behaviors of hash join operation on a relational database with both HDD and SSD environment, and clarify that it is necessary to consider the cost model by taking into ac characteristics of SSDs Key words DBMS, SSD, OLAP, Join 1. (HDD) CPU NAND flash Phase Change Memory FeRAM Storage Class Memory (SCM) HDD HDD 2 NAND flash-based Solid State Drive (SSD) NAND flash chip SSD HDD 1-1 HDD DBMS TPC-H [1] 1

2 SSD SSD I/O HDD DBMS 4. SSD DBMS SSD 2. 1 Grace [2] SELECT R. a2, S. a3 FROM R, S WHERE R. a1 = S. a1 R S R.a1 S.a1 R.a2 S.a3 R S build phase probe phase 2 phase build phase R S S probe phase S R R S 2. 2 [3] build phase n n S 1 R build phase R 1 S 1 probing S R HDD DBMS DBMS Grace Time[s] k 256k 1M 4M 16M 64M 256M 1G 4G working memory[byte] idle iowait sys usr HDD work mem (Query2) Fig. 1 Query Execution Time for each work mem size on HDD (Query2) 1 Table 1 HDD Experimental Platform of HDD CPU Xeon 2.4GHz x 2 L3 Cache Main Memory Storage 2 Table 2 8MB 24GB Hitachi HDS7211CLA332 x1 (RAID6 JCS VCRVAX-4F) Experimental Database Setup DBMS PostgreSQL [4] Shared buffer work mem 8GB 64KB - 2GB TPC-H Scale factor 1 1 HDD DBMS TPC-H work mem I/O wait I/O DBMS HDD SSD 1-1 read/write / I/O HDD SSD 3. SSD HDD 2

3 3 SSD Table 3 Experimental Platform of SSD CPU Xeon 2.27GHz x 4 L3 Cache 24MB Main Memory 64GB Storage iodrive Duo x4 (8LU, Software RAID) 4 Table 4 Table sizes and number of tuples Table name Size # of tuple part 3.2GB 2,, lineitem 86GB 6,37,92 orders 2GB 15,, customer 2.7GB 15,, supplier 173MB 1,, nation 8kB 25 region 8kB 5 5 Table 5 Latency of CPU caches and storages Storage Latency L3 cache 2 ns DRAM (local) 1 ns DRAM (Remote) 15 ns SSD (512KB) 3 us HDD (random, 512KB) 1 ms 3. SDD HDD 2 DB TPC-H 4 (work mem) 64KB-2GB perf Query1 lineitem part Query 1 Join line item and part SELECT ( ) FROM l i n e i t e m, part WHERE l p a r t k e y = p partkey Time[s] Fig idle iowait sys usr 64k 256k 1M 4M 16M 64M 256M 1G 4G 1.4e+1 1.2e+1 working memory[byte] work mem (Query1) Query Execution Time for each work mem size (Query1) 1e+1 8e+9 6e+9 4e+9 2e+9 cache-references cache-misses cache-miss-rate 64k 256k 1M 4M 16M 64M256M 1G 4G working memory [byte] rate[%] work mem L3 / (Query1) Fig. 3 Number of s/misses and rates for each work mem size (Query1) 2 work mem user system iowait idle CPU usr CPU sys iowait I/O 1 work mem 16MB 32MB L3 CPU 32MB L3 3 work mem L3 L3 32MB 5 1 iodrive I/O sys sys 3

4 work mem=4mb 1(ns) = 2(s) work mem=1gb 1(ns) = 9(s) 7 2 CPU work mem=64kb-512mb 4(a)-4(e) 4(a) 3 / build phase probe phase probe phase work mem L3 L3 work mem=4mb 6% work mem=16mb 1% work mem L3 work mem=32mb 4% work mem=512mb 8% build phase work mem build 1 probing work mem probing probe phase work mem=4mb 3% work mem=512mb 45% work mem=1gb 7% work mem=1gb 1 probe phase build phase build phase probe phase work mem L3 CPU PostgreSQL 1 1 probing lineitem = = CPU I/O 2 4MB 1GB I/O CPU CPU I/O CPU 5 Query2 Fig. 5 Execution Plan of Query TPC-H 8 Query2 5 Query 2 TPC-H Query 8 (join part) SELECT e x t r a c t ( year from o o r d e r d a t e ), l e x t e n d e d p r i c e (1 l d i s c o u n t ), n2. n name FROM part, s u p p l i e r, l i n e i t e m, orders, customer, nation n1, nation n2, r e g i o n WHERE p partkey = l p a r t k e y and s suppkey = l suppkey and l o r d e r k e y = o o r d e r k e y and o c u s t k e y = c c u s t k e y and c n a t i o n k e y = n1. n nationkey and n1. n r e g i o n k e y = r r e g i o n k e y and r name = AMERICA and s n a t i o n k e y = n2. n nationkey and o o r d e r d a t e between and and p t y p e = ECONOMY ANODIZED STEEL 6 work mem 1 L3 CPU work mem=2mb 431 work mem=64mb work mem L3 work mem=2mb 3% work mem=512mb 5% L3 8MB 5 3MB 4

5 1.2e+8 1e+8 8e+7 6e+7 4e+7 2e+7 1.2e+8 1e+8 8e+7 6e+7 4e+7 2e+7 rate build phase probe phase elapsed time[s] (a) work mem=4mb rate e+8 1e+8 8e+7 6e+7 4e+7 2e+7 1.2e+8 1e+8 8e+7 6e+7 4e+7 2e+7 rate (b) work mem=16mb rate e+8 1e+8 8e+7 6e+7 4e+7 2e+7 1.2e+8 1e+8 8e+7 6e+7 4e+7 2e+7 rate (c) work mem=32mb rate (d) work mem=128mb (e) work mem=512mb (f) work mem=1gb 4 L3 / (Query1) Fig. 4 Timeline of number of s/misses and rates (Query1) Time[s] Fig idle iowait sys usr 64k 256k 1M 4M 16M 64M 256M 1G 4G working memory[byte] work mem (Query2) Query Execution Time for each work mem size (Query2) work mem L3 5% 1 lineitem l orderkey 4 2 l orderkey probing 2 2 TPC-H Test DataBase Data Generation 7 1.2e+1 1e+1 8e+9 6e+9 4e+9 2e+9 cache-references cache-misses cache-miss-rate 64k 256k 1M 4M 16M 64M256M 1G 4G working memory [byte] rate[%] work mem L3 / (Query2) Fig. 7 Number of s/misses and rates for each work mem size (Query2) SSD I/O CPU 1 HDD 2 6 SSD SSD CPU HDD 5

6 DB HDD SSD SSD CPU 4. DBMS SSD HDD SSD Bhattacharjee temperature-aware caching (TAC) schema [5], [6] Kang FaCE [7] multiversion FIFO Second Chance SSD HDD Koltsidas [8] read-intensive SSD write-intensive HDD hstorage-db [9] DBMS semantic information FD-tree [1] SSD PIO B-tree [7] SSD internal parallelism SSD SSD SSD SSD CPU HDD SSD SSD SSD DBMS [1] Transaction Processing Performance Council, an ad-hoc, decision support benchmark. [2] M. Kitsuregawa, H. Tanaka, and T. Moto-Oka, Relational algebra machine grace, Proceedings of RIMS Symposium on Software Science and Engineering, pp , Springer- Verlag, London, UK, UK, [3] D.A. Schneider and D.J. DeWitt, A performance evaluation of four parallel join algorithms in a shared-nothing multiprocessor environment, Proceedings of the 1989 ACM SIGMOD international conference on Management of data, pp , SIGMOD 89, ACM, New York, NY, USA, [4] PostgreSQL. [5] B. Bhattacharjee, K.A. Ross, C. Lang, G.A. Mihaila, and M. Banikazemi, Enhancing recovery using an ssd buffer pool extension, Proceedings of the Seventh International Workshop on Data Management on New Hardware, pp.1 16, DaMoN 11, ACM, New York, NY, USA, 211. [6] M. Canim, G.A. Mihaila, B. Bhattacharjee, K.A. Ross, and C.A. Lang, Ssd bufferpool extensions for database systems, Proc. VLDB Endow., vol.3, no.1-2, pp , Sept. 21. [7] W.-H. Kang, S.-W. Lee, and B. Moon, Flash-based extended cache for higher throughput and faster recovery, Proc. VLDB Endow., vol.5, no.11, pp , July 212. [8] I. Koltsidas and S.D. Viglas, Flashing up the storage layer, Proc. VLDB Endow., vol.1, no.1, pp , Aug. 28. [9] T. Luo, R. Lee, M. Mesnier, F. Chen, and X. Zhang, hstorage-db: heterogeneity-aware data management to exploit the full capability of hybrid storage systems, Proc. VLDB Endow., vol.5, no.1, pp , June 212. [1] Y. Li, B. He, R.J. Yang, Q. Luo, and K. Yi, Tree indexing on solid state drives, Proc. VLDB Endow., vol.3, no.1-2, pp , Sept DBMS SSD 1 L3 CPU SSD I/O CPU CPU HDD DB 6