Kaminario climbs performance mountain
K2 DRAM grid-in-a-box
SaaS data loss: The problem you didn’t know you had
A start-up called Kaminario has a new spin on in-memory datasets with its K2 DRAM-based grid of storage nodes having no single point of failure, unlike Texas Memory Systems, it says, and being cheaper than flash-based million-IOPS products.
The idea is an old one: hold application data sets in externally-attached, solid-state storage and get radically faster access to data. There have been two main approaches. One is to use DRAM and the main exponent of this is Texas Memory Systems (TMS) with its RamSan set of products. The other is to use flash, and Fusion-io has been making the running here with one million IOPS demos with IBM via its SAN Volume Controller, and another with HP, both using Fusion's ioDrive flash memory card. IBM is now using STEC solid state drives (SSDs) in an SAN Volume Controller flash-based product.
Kaminario's concept is to get its DRAM by using commercial, off-the-shelf hardware - Dell blade servers in this case. These are combined into a grid using a 1GigE interconnect and connected to the outside world by a minimum of two clustered I/O Directors with two 8Gbit/s Fibre Channel ports each. Host servers see a standard block device.
The I/O Directors control access to a set of Data Nodes, the Dell blades, whose operation is coordinated by KOS operating system software. Each data node consists of its CPU, DRAM and two hard drives. Application data is written across Data Nodes. Primary Data nodes hold the data in DRAM with a mirror in one in one of its hard drives. A Secondary Data Node holds a backup copy in its hard drive. Should a node fail then data can be reconstructed.
Reads are executed in parallel from all data nodes which hold the data set referred to. The K2 appliance has full redundancy for all components, with Kaminario pointing out that TMS RamSan products do have a single point of failure. (See discussion here.)
Kaminario criticises flash-based alternatives to K2, saying that useful life is less than DRAM and block erase-write-cycles reduce performance, as do other write-based processes.
The minimum K2 configuration is two I/O Directors and four Data Nodes offering 500GB raw DRAM capacity, 300,000 random IOPS and 3.2GB/sec bandwidth. The maximum supported configuration is ten Data Nodes providing 1.5 million IOPS and 16GB/sec bandwidth. Customers can start small and buy more I/O Directors for performance and more Data Nodes for capacity.
The entry-level configuration seems to cost around $200,000. Kaminario is coy about its pricing, which is based on a $ per GB and IOPS calculation.
Which other suppliers' products can do a million IOPS from solid state? Fusion-io did it with IBM's SVC using 4TB of flash. HP did it with 2.5TB of Fusion-io flash. LSI has reported getting a million IOPS from 1TB of Seagate Pulsar flash. TMS did the same with its RamSan 5000 in October 2008 - actually ten RamSan 500s with 640G of DRAM cache in total and an undisclosed amount of flash. A RamSan 500 can hold up to 2TB of single level cell (SLC) flash but the idea of a 20TB system generating just one million IOPS seems bizarre.
Kaminario claims it offers ultra high performance. TMS has a RamSan-630 offering 4GB/sec throughput, 10TB of capacity and 500,000 sustained IOPS from its SLC flash. Its RamSan-440 has up to 512GB of DRAM, up to 660,000 IOPS and 4.5GB/sec throughput.
Violin Memory's latest 3200 product has from 500GB to 10TB of SLC flash and delivers a claimed 220,000 sustained write IOPS. VION has a HyperStore-6200 100TB DRAM box, using TMS RamSan 620 technology, with five million IOPS and 60GB/sec bandwidth.
Discounting the VION monster, the K2 would appear to have both an IOPS edge and a throughput advantage. Kaminario makes a big thing of the K2 being more reliable than TMS or other suppliers' products but how many RamSan's have failed in operation? Kaminario also says its product is an appliance and can be installed and running in a day with no changes needed to applications or the overall environment. ®
COMMENTS
Again missing a few facts
The TMS RamSans have duplicate controllers internally. The internal components are RAIDed at the chip level as well as having ECC and chipKill technologies . We get redundancy for the backplane the same way Kaminario does, by having more than one system. For customers that require that full redundancy we sell them two that they use RAID 1 to mirror. RamSans are configured using an easy to use web based GUI interface. From box to being used takes about 30 minutes.
Our RamSan440 is DDR RAM based with up to 512 gigabytes of fully usable capacity that provides 600,000 IOPS at 15 microsecond latency. We can provide 1,200,000 IOPS in 6U of space using the RamSan440. Two RamSan630s, again in 6u, can give you 1,000,000 IOPS at 80 microsecond write and 250 microsecond read times if you don't like the price on the RamSan440 but still want the IOPS.
All our numbers are backed up by published SPC-1 results. I don't see any certified published results from Kaminario, just a bit of hand waving. Afraid Kaminario has a bit more mountain to go.
Mike Ault
(Mike works for TMS)
IT infrastructure monitoring strategies
Agentless Backup is Not a Myth
Top 10 SIEM implementer’s checklist
Steps to Take Before Choosing a Business Continuity Partner
Enabling efficient data center monitoring
