imcsummit 2016 keynote - benzi galili - more memory for in-memory easy

10
MORE MEMORY FOR IN-MEMORY? EASY! BENZI GALILI CHIEF OPERATING OFFICER SCALEMP, INC. See all the presentations from the In- Memory Computing Summit at http://imcsummit.org

Upload: in-memory-computing-summit

Post on 21-Jan-2017

54 views

Category:

Data & Analytics


0 download

TRANSCRIPT

MORE MEMORY FOR IN-MEMORY? EASY!BENZI GALILICHIEF OPERATING OFFICERSCALEMP, INC.

See all the presentations from the In-Memory Computing Summit at http://imcsummit.org

IN MEMORY IS STRONG AND GROWINGVertical scaling: make it a

realityWhen you ask for 30TB memory you are no

longer considered insane

Horizontal scaling: fewer nodes

You already use multiple nodes, but more memory per node enables more processing

STATE OF THE INDUSTRY

Vertical scaling has a glass ceiling Commodity 2S server max out at 768GB, or 1.5TB for much more $$ High-end systems reach 6TB-12TB for hefty premiums “Large-scale” cloud instance maxes out much lower (Amazon 244GB EC2 @ $4.2/GB/month!)

Horizontal scaling is challenging Not all applications can be (re)written as distributed The more nodes being used, the higher the complexity, and

cost

The alternatives aren’t really “in-memory” Programming for flash (e.g. aerospike) is not as-trivial as in-

memory SCM requires explicit programming to overcome latency gaps

SDM-S

SDM-F

SDC

Software-defined Computing Aggregate all computing resources of multiple physical

servers into a single virtual SMP system using high-performance fabric

Supercomputing, Scientific and Engineering applications Up to 32k CPUs and 2PB RAM in a single system image

Software-defined Memory / Fabric Leverage the memory of other nodes to create very

large system memory, operating at DRAM performance Large in-memory databases and in-memory computing Up to 2PB RAM

Software-defined Memory / Storage Leverage high-performing NVM to create large

economical system memory in a single server In memory computing at scale, lower TCO Up to 128TB system memory Future proof: 3D Xpoint, ReRam, NVDIMM, etc.

NVMeNVMe NVMe NVMe NVMe NVMe

SOFTWARE DEFINED: COMPUTING AND MEMORY

HOW GOOD IS SDM-F?In-Memory Commercial Database using 6TB

memory

8 Sockets6 TB RAM

Local DRAM

8 Sockets1.5 TB RAM

SDM-F2.25 TB 2.25 TB

SDM-F Memory Pool

100 %

98 %

In daily use with 30TB (6+24)

RAM

Data Processing Throughput

HOW GOOD IS SDM-S?

memcached using 500GB

512GB server w/RAM (100%)vs.

128GB server w/SWAP (48%)

HOW GOOD IS SDM-S?

memcached using 500GB

512GB server w/RAM (100%)vs.

128GB server w/SWAP (48%)vs.

128GB server w/SDM-S (98%)

Note: Same SSDs were used for SWAP and SDM-S

HOW GOOD IS SDM-S?

memcached using 500GB

512GB server w/RAM (100%)vs.

128GB server w/SWAP (48%)vs.

128GB server w/SDM-S (98%)vs.

64GB server w/SDM-S (80%)

Note: Same SSDs were used for SWAP and SDM-S

1:4 ratio

1:8 ratio

1:4 ratio

SUMMARY Software-defined memory is the solution for in-memory scaling

Turns vertical scaling (scale-up) into an affordable reality Enables dramatically lower TCO for horizontally scaled (scale-out) applications

ScaleMP provides a holistic high-end-virtualization solution: SDC: virtually infinite server scaling, on-demand, shipping since 2005 SDM-F: “any large memory need”, high-performance, up to 2PB of RAM, shipping

since 2012 SDM-S: high-concurrency workloads, unbelievable economics, up to 128TB of

memory

SDM-S general availability by key NVM vendors in 2016! Early access deployments at select customers

THANK [email protected]