• 12/24/2013
    10:00 AM
  • Rating: 
    0 votes
    Vote up!
    Vote down!

Amazon Boosts High-Speed I/O Instances With SSDs

Amazon Web Services follows Rackspace, Colt, and Digital Ocean in ramping up high I/O servers with solid state disks.
Top 10 Cloud Fiascos
Top 10 Cloud Fiascos
(click image for larger view)

Amazon Web Services introduced I2, its second generation of high I/O instance types, Monday, after the first generation, H1, proved too bulky for some users seeking a server optimized for random I/Os.

Not every data-capture or call for stored data amounts to a sequential read or write, something that spinning disks are good at. On the contrary, many are for random reads or writes, something that solid state disks (SSDs) are good at. Spinning hard drives, on the other hand, can prove slow at random data retrieval as the head moves mechanically to the proper area of the spinning plate. The four new I2 instances are equipped with SSDs to eliminate that latency.

Amazon is somewhat late in coming to the solid state party. In Europe, Colt implemented extensive SSD-based services earlier, as did Rackspace and startup DigitalOcean in the US.

Amazon is turning to SSDs to try to better equip the random-access crowd wanting to do unpredictable relational database reads or transactional data writes that may represent small amounts of data, but require fast access to unpredictable areas of storage.

"I2 instances are well-suited for transactional systems, high-performance NoSQL databases like Cassandra and MongoDB, and other applications that benefit from very high disk I/O performance," AWS's announcement said.

[Want to learn more about Amazon EC2 compute instances? See Amazon Cuts Some M3 Compute Instances 10%.]

The I2s are expected to replace the HI1 instances, previously designated as the instances suitable for high I/O server types. "Customers (using HI1s) have told us that they not only wanted even more high I/O performance, but also wanted smaller instance size options for their smaller database fleets," said Matt Garman, VP of Amazon EC2. The I2 instances have been given twice the memory of roughly equivalent HI1 instances, he said in the announcement.

The four new I2 instances are part of Amazon's storage-optimized family of instance types. Equipped with SSDs, they speed random I/O throughput. The I2s are capable of 365,000 random read I/O operations per second and 315,000 random write I/O ops per second. The I2s achieve such throughput rates based on hosts using the latest Intel Ivy Bridge processors, also known as Xeon E5-2670 version 2 CPUs.

The I2s are offered in four types:

The smallest is the I2 extra-large, with a processor consisting of four virtual CPUs equal to 14 EC2 Compute Units (ECUs). (An ECU is the equivalent of a 2007 Intel Xeon core running at 1 GHz.) The I2.xlarge comes with 30.5 GBs of RAM and one 800-GB SSD. It is priced at 85.3 cents an hour.

The next instance type is the I2 double extra-large, with eight virtual CPUs or 27 ECUs, 61 GB of RAM, and two 800-GB SSDs. It is priced at $1.705 an hour.

There is also the I2 quadruple extra-large, with 16 virtual CPUs or 53 ECUs, 122 GB of RAM, and four 800-GB SSDs. It is priced $3.41 an hour.

The largest is the I2 8x extra-large, with 32 virtual CPUs or 104 ECUs, 244 GB of RAM, and eight 800-GB SSDs. It is available at $6.82 per hour.

"We host 180,000 applications on Parse using MongoDB and Cassandra, and these applications need high memory and high IOPS. We have been eagerly awaiting I2 instances," said Charity Majors, operations tech lead at, in the announcement. Parse is a mobile application development platform.

Not included in the announcement, but listed in the same storage-optimized family, is a fifth instance type, one geared for more sequential read/writes, the HS1.8xtralarge. The HS1 has access to 24 2-TB local hard drives, rather than SSDs. It is equipped with a virtual CPU equivalent to 35 ECUs, which falls in between the I2 double extra-large and I2 quadruple extra-large in server size. Its 117 GB of RAM is close in size to the quadruple extra-large I2. But its overall storage capacity is much larger. While the I2 has 3.2 TB of SSDs, the HS1.8xlarge has 24 2-terabyte local hard drives. Amazon terminology is not always precisely defined in its announcements, but local appears to mean access to hard drives on the same server cluster.

The HS1.8xtralarge is priced at $4.60 per hour.

The current generation HS1 instance can be used to illustrate the growing power of the instance types for the price charged. The $4.60 per hour contrasts with what Amazon lists as the previous generation's HI1 quadruple extra-large, which had similar CPU, half the memory, and 1/24th the storage for $3.10 an hour.

Charles Babcock is an editor-at-large for InformationWeek, having joined the publication in 2003. He is the former editor-in-chief of Digital News, former software editor of Computerworld, and former technology editor of Interactive Week.

Private clouds are moving rapidly from concept to production. But some fears about expertise and integration still linger. Also in the Private Clouds Step Up issue of InformationWeek: The public cloud and the steam engine have more in common than you might think. (Free registration required.)


SSDs and cloud: meant for each other

Solid state disks are extremely popular with cloud users for their ability to reduce load times, speed response times. With their known failure rates, they need to be managed by the cloud software that can anticipate hardware failures. But that's what a well-architected cloud is designed to do anyway, with or without SSDs.

Re: SSDs and cloud: meant for each other

Can you see this choice being abstracted away? I think in the long run what you'd want are intelligent storage systems that choose the correct media automatically based on the needs of an application.

Naturally, you'd always want the option of making that choice yourself. But shouldn't cloud operating systems start to do optimizations of this sort more proactively?

Re: SSDs and cloud: meant for each other

That's nuts. You're either talking about tens of Gigs (or more) being instantly read/written to/from either SSD or cheap drives.  You going to magically migrate that data on demand?  Or mirror it (at different speeds)? At that point.. what's the point?

Probably the BEST way to do it is to separate high speed filesystem space from low-speed filesystem space and offer customized preconfigured setups. Dynamic gear shifting just doens't make sense here



Are SSDs commodity parts yet?

That's a good point, David, this could be done proactively by the cloud service. But so far, SSDs have been too pricey a component to make it into many cloud architectures. First they appeared as a paid for service, where the use had the option of selecting a quality of service level for his I/O operations. Then DigitalOcean and others incorporated them into basic cloud servers as a standard feature, without extra charge. For the time being, cloud providers will keep services bare bones and go for the maximum number of users at the lowest possible price. But they could evolve more and more sophisticated I/O ops and storage system choices. But SSDs will have to become commodity parts for that to happen on a large scale.

Re: Are SSDs commodity parts yet?

If we talk about SSD, A typical SSD uses what is called NAND-based flash memory, but one question which i guess is most common as well, What does non-volatile mean ?

Re: Aditshar and non-volatile

Re: Aditshar. "Non-volatile" means the data won't be lost when the power is shut off. The NAND-based SSDs manufactured today are non-volatile. Solid state memory used for RAM is typically built from chips that are volatile. The data disappears and can't be recovered when the power disappears, although some manufacturers build solid state devices using volatile memory chips, then include a backup supply of power in the event of power loss through a battery. It only takes a trickle of power to maintain the data, so the battery serves well as long as the outage is short.