Memory and storage units
From HP-SEE Wiki
Contents |
Memory technologies
Phase Change Memory (PCM)
One of the more promising solid state memory technologies on the horizon is Phase Change Memory (PCM). PCM has the potential to write data faster than current DRAM chips by charging atoms within a crystal, and this can led that the technology might enable instantaneous computer boot-up.
At the atomic level, PCM stores data in a compound of germanium, antimony and tellurium. When a voltage is applied to the atoms, they change into an ordered crystal. The data can then be deleted by melting the crystalline substance. To read the information, a computer determines the electrical resistance of the material. An important attribute of phase change memory is that the technology is non-volatile. This means it does not require power to retain information like standard RAM offerings. Along with the possibility of replacing system memory, these chips might end up competing with NAND flash as well. Despite the benefits, PCM suffers from an inherent issue that has slowed its path to adoption. The biggest one is its write speed. Current DRAM technology can perform write operations within a 1-10 nanosecond window, which is faster than the time it takes for the germanium, antimony and tellurium compound in PCM to crystallize. Other crystalline compounds with faster reaction times have been researched, but they are not as stable as the current PCM design, slowly erasing themselves in low temperatures over time.
Magnetoresistive random-access memory (MRAM)
Magnetoresistive random-access memory (MRAM) is a non-volatile random-access memory technology currently being developed by several vendors such as IBM, HP and others.
MRAM is a method of storing data bits using magnetic charges instead of the electrical charges used by DRAM (dynamic random access memory). Scientists define a metal as magnetoresistive if it shows a slight change in electrical resistance when placed in a magnetic field. By combining the high speed of static RAM and the high density of DRAM, proponents say MRAM could be used to significantly improve electronic products by storing greater amounts of data, enabling it to be accessed faster while consuming less battery power than existing electronic memory.
Stacked memory
Since the memory wall (in terms of performance and efficiency) has been the limiting factor of HPC systems for many years, three-dimensional (3D) die-stacking has received a great attention in the HPC community. 3D stacking enables the construction of circuits using multiple layers of active silicon bonded with low-latency, high-bandwidth and very dense vertical interconnects. Also, 3D integration would enable stacking memory directly on top of a microprocessor, resulting in the big reduce of wire delay between the two. Many studies have been conducted in the scientific community that shown viability of such a solution, so it attracted memory vendors to offer commercial products based on the idea.
Companies that announced products based on 3D stacked memory include Samsung, Micron, NVIDIA, Intel and others.
Links:
Hybrid Memory Cube
Hybrid Memory Cube (HMC) is a new memory architecture proposed by HMC Consortium led by Micron, Samsung, Intel, IBM, Microsoft, ARM and others. HMC chnages the way in how memory is used in the system by tightly coupling intelligent memory with CPUs, GPUs, and ASICs, thus allowing systems to enable dramatic improvements in efficiency and power optimization.
At the core of the HMC is a small, high-speed logic layer that sits below vertical stacks of DRAM die that are connected using through-silicon-via (TSV) interconnects. The DRAM has been designed solely to handle data, with the logic layer handling all DRAM control within the HMC. HMC could be mounted directly adjacent to the processors for best performance, or in a scalable module form factor for optimized power efficiency.
Hybrid Memory Cube is a a hybrid of memory and logic that comes in either four- or eight-layer stacks of memory. The logic layer is a memory controller that will work like a hypervisor for testing, routing and optimization. The new device is a true 3D stack, including between 2,000 and 3,000 through-silicon vias. The die themselves will be manufactured at the 20nm process node or smaller, with an expected jump in throughput that will enable movement of the same amount of data for approximately 70% less power.
HMC advantages include:
- Reduced latency - More responders built into HMC will lower queue delays and allow higher bank availability.
- Increased bandwidth - A single HMC can provide more than 15x the performance of a DDR3 module.
- Power reductions — HMC uses 70% less energy per bit than DDR3.
- Smaller physical systems — Stacked architecture of HMC uses nearly 90% less space than today’s RDIMMs.
- Adaptability to multiple platforms — Logic-layer flexibility allows HMC to be tailored to multiple platforms and applications.
First modules are expected in the second half of 2014.
Links:
- http://www.micron.com/products/hybrid-memory-cube
- http://www.hpcwire.com/hpcwire/2012-07-10/hybrid_memory_cube_angles_for_exascale.html
- http://lp-hp.com/blog/2011/10/06/samsung-micron-unveil-3d-stacked-memory-and-logic/
Crossbar memory
Crossbar is resistive RAM technology that can store up to 40x more data than other contemporary memory technologies. It is believed to be the most compact memory available today with the added benefit of being significantly more energy efficient than both flash and DRAM. RRAM-based memory technology can scale up to 1 TB and can be incorporated into any standard CMOS manufacturing fabrication process. The technology is capable of 3D stacking adn enables the integration of multiple memory array layers on top of the CMOS controller, thus providing potential 1TB storage capacity.
Storage technologies
Fujitsu-Eternus-DX90S2-ECO
Power consumption reduction - Eco-mode
NIIFI has several DX90S2 disk storage systems at remote sites spread across Hungary. These storage devices have a feature called Eco-mode that could help lowering power consumption and thus also reducing environmental burdens.
Read more here: http://www.fujitsu.com/global/services/computing/storage/eternus/products/diskstorage/feature/strsys-g01.html
ECO functions in DX90S2
There are functions to
- set ECO mode
- create/modify ECO schedules
- set ECO schedule for a RAID group
---
ECO functions cannot be used unless the ECO mode is enabled.
Q: How to enable ECO mode?
A:
set eco-mode [-mode {enable | disable}] [-monitoring-time {10 | 20 | 30 | 40 | 50 | 60}] [-limit {1 | 2 | 3 | ... | 25}]
Monitoring time refers to the monitoring interval time of the host I/O. Limit refers to the maximum number of times that the disk motor will power off and on in one day. Both monitoring time and limit are optional. If omitted, the existing setting is not changed.
---
"An ECO schedule is a schedule that applies ECO functions. The ECO schedule specifies the time that a disk motor is rotated (it does NOT specify when the motor will be turned off). A maximum of 64 ECO schedules can be created. One ECO schedule consists of a maximum of 8 ECO schedule events (at least one ECO schedule event is required in one ECO schedule)."
ECO schedules are created to fit your specific needs. Best practice would be monitoring your storage resources, gathering usage stats, analyzing access patterns and then create ECO schedules according to those information. As user behaviour may change over time you should then continue monitoring and occasionally validate if your schedules are still appropriate.
Q: How to create ECO schedule?
A:
create eco-schedule -name schedule_name -event-type {every-day | every-week,{mon | tue | wed | thu | fri | sat | sun},{mon | tue | wed | thu | fri | sat | sun} | specific-day,MMDD,R | specific-week,MM,{1st | 2nd | 3rd | 4th | last},{mon | tue | wed | thu | fri | sat | sun},{mon | tue | wed | thu | fri | sat | sun} } -event-from hhmm -event-to hhmm
'Event from' refers to a specific starting time for the schedule. Format is "hhmm". "hh" is the starting hour (00-23) and "mm" is the starting minute (00 or 30). A disk motor will power ON at the time specified by this parameter.
'Event to' refers to a specific end time for the schedule. The format is "hhmm". "hh" is the ending hour (00-23) and "mm" is the ending minute (00 or 30).
"This command can create one ECO schedule and only one ECO schedule event. If two or more ECO schedule events are created and defined, use the 'set eco-schedule' command."
Examples:
create eco-schedule -name SC001 -event-type every-day -event-from 0600 -event-to 1800
An ECO schedule 'SC001' is created. It will apply every day from 06:00 to 18:00.
create eco-schedule -name SC002 -event-type every-week,tue,fri -event-from 0600 -event-to 1800
This schedule will apply every week from Tuesday to Friday from 06:00 to 18:00.
create eco-schedule -name SC003 -event-type specific-day,0501,3
This schedule will apply for three days, from May 1st to May 3rd.
create eco-schedule -name SC004 -event-type specific-week,10,2nd,mon,wed
This schedule will apply from the second Monday to the second Wednesday in October.
Q: How to change ECO schedule or add ECO events?
A:
set eco-schedule {-schedule-number schedule_number | -schedule-name schedule_name} [-name schedule_name] [-event-number {1 | 2 | 3 | 4 | 5 | 6 |7 | 8}] [-event-type {every-day | every-week,{mon | tue | wed | thu | fri | sat | sun}, {mon | tue | wed | thu | fri | sat | sun} | specific-day,MMDD,R | specific-week,MM,{1st | 2nd | 3rd | 4th | last}, {mon | tue | wed | thu | fri | sat | sun},{mon | tue | wed | thu | fri | sat | sun}}] [-event-from hhmm] [-event-to hhmm]
Example:
set eco-schedule -schedule-name SC001 -event-number 2 -event-type every-day -event-from 0600 -event-to 1800
Adds ECO schedule event #2 in the ECO schedule named 'SC001'.
Q: How to delete ECO schedules?
A:
delete eco-schedule {-schedule-number schedule_numbers | -schedule-name schedule_names} [-event-number {1 | 2 | 3 | 4 | 5 | 6 | 7 | 8}]
Examples:
delete eco-schedule -schedule-name SC002
Delete the ECO schedule named 'SC002'.
delete eco-schedule -schedule-name SC001 -event-number 2
Delete ECO schedule event #2 in the ECO schedule named 'SC001'.
Note that only one event number can be specified.
If the event-number is omitted, the ECO schedule is deleted.
delete eco-schedule -schedule-name SC002,SC003
Delete both the ECO schedule named 'SC002' and 'SC003' at the same time.
---
Q: How to use ECO schedules?
A: ECO schedules are associated with RAID groups.
Note that ECO mode cannot be used for RAID groups that contain SSDs.
Q: How to associate an ECO schedule with a RAID group?
A:
set eco-raid-group {-rg-number rg_numbers | -rg-name rg_names} [-schedule-number schedule_number | -schedule-name schedule_name] -action {enable | disable}
Both RAID group and schelude are identified by their number or name. 'Action' specifies whether the associated definition is enabled and thus it could be 'enable' or 'disable'.
Examples:
set eco-raid-group -rg-name RGP001 -schedule-name SC001 -action enable
Associates the ECO schedule named 'SC001' with the RAID group named 'RGP001'.
set eco-raid-group -rg-number 1-10 -schedule-name SC001 -action enable
Associates the ECO schedule named 'SC001' with the RAID groups #1-10 at the same time.
set eco-raid-group -rg-name RGP001 -action disable
Disables the ECO function of the RAID group named 'RGP001'.
set eco-raid-group -rg-name RGP001 -schedule-number 1 -action enable set eco-raid-group -rg-name RGP001 -schedule-number 2 -action enable
If you associate ECO schedule #1 with the RAID group named 'RGP001' and then associate the ECO schedule #2 with the same RAID group then the result is ECO schedule #1 being released as well as ECO schedule #2 associated with 'RGP001'.
Q: How to release ECO schedule associated with a specific RAID group? A: release eco-raid-group {-rg-number rg_numbers | -rg-name rg_names}
Examples:
release eco-raid-group -rg-name RGP001
Releases an ECO schedule associated with the RAID group named 'RGP001'.
release eco-raid-group -rg-number 1-10
Releases an ECO schedule associated with RAID groups #1 - #10 at the same time.
---
Documentation and examples are found in "Eternus CLI User's Guide - Eternus DX80 S2/DX90 S2, Eternus DX410 S2/DX440 S2".
DDN EXAScaler
DDN EXAScaler, is a next generation parallel file storage system appliance that combines DDN HPC storage technology with the open-source Lustre File System.
HPC applications that require massive throughput and highly parallel access are challenging traditional SAN and NAS storage technologies. Storage systems designed to serve the needs of the last 10 years are no longer able to deliver the scalable bandwidth and file system locking capabilities required by today's HPC environments. To satisfy these requirements, an extremely high performance approach which integrates file system and storage media into one extremely high performance file storage appliance is needed.
