I'm glad this is the final day of the IBM Systems Technical Conference (STC08) here in Los Angeles.While I enjoyed the conference, one quickly reaches saturation point with all the information presented.
- XIV Architecture Overview
Before this conference, many of the attendees didn't understandIBM's strategy, didn't understand Web 2.0 and Digital archive workloads,and didn't understand why IBM acquired XIV to offer "yet another disk systemthat servers LUNs to distributed server platforms." Brian Shermanchanged all that!
Brian Sherman, IBM Advanced Technical Support (ATS), is part of the exclusive dedicated XIVtechnical team to install these boxes at client locations, so he is very knowledgeable with the technical aspects of the architecture. He presented what the current XIV-branded model that clients can purchase now in select countries, and what the IBM-branded model will change when available worldwide.
Those who missed my earlier series on XIV can find them here:
Beyond this, Brian gave additional information on how thin provisioning, storage pools, disk mirroring, consistency groups, management consoles, and microcode updates are implemented.
- N series and VMware Deep Dive
Norm Bogard, IBM Advanced Technical Support, presented why the IBM N series makes such great disk storage for VMware
deployments. This wasclearly labeled as a "deep dive", so anyone who got lost in all of theacronyms could not blame Norm for misrepresentation.
IBM has been doing server virtualization for over 40 years, so it makes sense thatit happens to be the number one reseller of VMware offerings.VMware ESX server is a hypervisor that runs on x86 host, and provides an emulationlayer for "guest Operating Systems". Each guest can hvae one or more virtualdisks, which are represented by VMware as VMDK files. VMware ESX server acceptsread/write requests from the guests, and forwards them on to physical storage.Many of VMware's most exciting features requires storage to be external to thehost machine. [VMotion]allows guests to move from one host to another, [Distributed Resource Scheduler (DRS)]allows a set of hosts to load-balance the guestsacross the hosts, and [High Availability (HA)] allows the guests on a failed hostto be resurrected on a surviving host. All of these require external disk storage.
ESX server allows up to 256 LUNs, attached via FCP and/or iSCSI, and up to 32 NFS mount points. Across LUNs, ESX server uses VMFS file system, which is a clusteredfile system like IBM GPFS that allows multiple hosts to access the same LUNs.ESX server has its own built-in native multipathing driver, and even provides FCP-iSCSIand iSCSI multipathing. In other words, you can have a LUN on an IBM System Storage N series thatis attached over both FCP and iSCSI, so if the SAN switch or HBA fails, ESX servercan failover to the iSCSI connection.
ESX server can use NFS protocol to access the VMDK files instead. While the default is only 8 NFS mount points, you can increase this to 32 mount points. NAS can takeadvantage of Link Aggregate Control Protocol [LACP] groups, what some call "trunking" or "EtherChannel". This is the ability to consolidate multiple streams onto fewer inter-switch Ethernet links, similar to what happens on SAN switches.For the IBM N series, IBM recommends a "fixed" path policy, rather than "most recently used".
IBM recommends disabling SnapShot schedules, and setting the Snap reserve to 0 percent.Why? A snapshot of an ESX server datastore has the VMDK files of many guests, all of which would have had to quiesce or stop to make the data "crash consistent" for theSnapshot of the datastore to even make any sense. So, if you want to take Snapshots, itshould be something you coordinate with the ESX server and its guest OS images, and notscheduled by the N series itself.
If you are running NFS protocol to N series, you can turn off the "accesstime" updates. In normal file systems, when you read a file, it updates the"access time" in the file directory. This can be useful if you are looking forfiles that haven't been read in a while, such as software that migrates infrequentlyaccessed files to tape. Assuming you are not doing that on your N series, you might as well turnoff this feature, and reduce the unnecessary write activity to the IBM N series box.
ESX server can also support "thin provisioning" on the IBM N series. There isa checkbox for "space reserved". Checked means "thick provisioning" and uncheckedmeans "thin provisioning". If you decide to use "thin provisioning" with VMware,you should consider setting AutoSize to automatically increase your datastorewhen needed, and to auto-delete-snap your oldest snapshots first.
The key advantage of using NFS rather than FCP or iSCSI is that it eliminates theuse of the VMFS file system. IBM N series has the WAFL file system instead, andso you don't have to worry about VMFS partition alignment issue. Most VMDK aremisaligned, so the performance is sub-optimal. If you can align each VMDK to a32KB or 64KB boundary (depending on guest OS), then you can get better performance.WAFL does this for you automatically, but VMFS does not. For Windows guests, use "Windows PE" to configurecorrectly-aligned disks. For UNIX or Linux guests, use "fdisk" utility.
- What Industry Analysts are saying about IBM
Vic Peltz gave a presentation highlighting the accolades from securities analysts, IT analysts, and newsagencies about IBM and IBM storage products. For example, analysts like that IBM offersmany of the exciting new technologies their clients are demanding, like "thin provisioning", RAID-6 double-drive protection,SATA and Solid State Disk (SSD) drive technology.Analysts also like that IBM is open to non-IBM heterogeneous environments. Whereas EMC Celerra gateways supportonly EMC disk, IBM N series gateways and IBM SAN Volume Controller support a mix of IBM and non-IBM equipment.
Analysts also like IBM's "datacenter-wide" approach to issues like security and "Green IT". Rather than focusingon these issues with individual point solutions, IBM attacks these challenges with a complete"end-to-end" solution approach. A typical 25,000 square foot data center consumes $2.6 million dollars USD in power andcooling today, and IBM has proven technologies to reduce this cost in half. IBM's DS8000 on average consume26.5 to 27.8 percent less electricity than a comparable EMC DMX-4 disk system. IBM's tape systemsconsume less energy than comparable Sun or HP models.
- IBM iDataPlex product technical presentation
Vallard Benincosa, IBM Technical Sales Specialist, presented the recently-announced [IBM System x iDataPlex].This is designed for our clients that have thousands of x86 servers, that buy servers "racks at a time", tosupport Web 2.0 and digital archive workloads. The iDataPlex is designed for efficient power and cooling,rapid scalability, and usable server density.
iDataPlex is such a radical design departure, that it might be difficult to describe in words.Most racks take up two floor tiles, each tile is 2 foot by 2 foot square. In that space, a traditionalrack would have servers that were 19 inches wide slide in horizontally, with flashing lights and hot-swappabledisks in the front, and all the power supply, fans and networking connections in the back. Even with IBM BladeCenter,you have chassis in these racks, and then servers slide in vertically in the front, and all of the power supply, fanand networking connections in the back. To access these racks, you have to be able to open the door on boththe front and back. And the cooling has to go through at least 26.5 inches from the front of the equipment to the back.
iDataPlex turns the rack sideways. Instead of two feet wide, and four feet deep, it is four feet wide, and two feet deep.This gives you two 19 inch columns to slide equipment into, and the air only has to travel 15 inches from frontto back. Less distance makes cooling more efficient.
Next, iDataPlex makes only thing in the back the power cord, controlled by an intelligent power distribution unit (iPDU) so you can turnthe power off without having to physically pull the plug. Everything else is serviced from the front door.This means that the back door can now be an optional "Rear Door Heat Exchanger" [RDHX] that is filled with running water to makecooling the rack extremely efficient. Water from a cooler distirubtion unit (CDU) can power about threeto four RDHX doors.
Let's say you wanted to compare traditional racks with iDataPlex for 84 servers. You can put 42 "1U" serversin two racks each, each rack requires 10 kVA (kilo-volt-amps) so you give it two 8.6 kVA feeds each, that is fourfeeds, and at $1500-2000 dollars USD per month, will cost you $6000-8000. The iDataPlex you can fit 84 serversin one 20 kVA rack, with only three 8.6 kVA feeds, saving you $1500-2000 dollars USD per month.
Fans are also improved. Fan efficiency is based on their diameter, so small fans in 1U servers aren't as effective as iDataPlex's 2U fans, saving about 12-49W per server. Whereas typical 1U server racks spend 10-20percent of their energy on the fans, the iDataPlex spends only about 1 percent, saving 8 to 36 kWH per year per rack.
Each 2U chassis snaps into a single power supply and a bank of 2U fans. A "Y"power cord allows you to have one cord for two power supplies. A chassis can hold either two small server "flexnodes"or one big "flexnode". An iDataPlex rack can hold up to 84 small servers or 42 big servers. Since each "Y" cord can power up to four "flexnode" servers, you greatly reduce the number of PDU sockets taken,leaving some sockets available for traditional 1U switches.
The small "flexnode" server can have one 3.5 inch HDD, or two 2.5 inch HDD, either SAS or SATA, and the big "flexnode" can have twice these.If you need more storage, there is a 2U chassis that holds five 3.5 inch HDD or eight 2.5 inch HDD. These areall "simple-swappable" (servers must be powered down to pull out the drives). For hot-swappable drives, a 3Uchassis with twelve 3.5 inch SAS or SATA drives.
The small "flexnode" server has one [PCI Express] slot, the big servers have two. Thesecould be used for [Myrinet] clustering. With only 25W power,the PCI Express slots cannot support graphics cards.
The iDataPlex is managed using the "Extreme Cluster Administration Toolkit" [XCAT]. This is an open source project under Eclipse that IBM contributes to.
Finally was the concept of "pitch". This is the distance from the center of one "cold aisle" to the next "cold aisle".On typical data centers, a pitch is 9 to 11 tiles. With the iDataPlex it is only three tiles when using the RDHX doors, or six tiles without. Most data centers run out of power and cooling before they run out of floor space, so having more dense equipmentdoesn't help if it doesn't also use less electricity.Since the iDataPlex uses 40 percent less power and cooling, you can pack more racks persquare foot of an existing data center floor with the existing power and cooling available. That is what IBM calls "usable density"!
- What Did You Say? Effective Questioning and Listening Techniques
Maria L. Anderson, IBM Human Resources Learning, gave this "professional development" talk. I deal with different clients every week, so I fully understand that there is a mix of art and science incrafting the right questions and listening to the responses.The focus was on howto ask better questions and improve the understanding and communication during consultative engagements. Thisinvolves the appropriate mix of closed and open-ended questions, exchanging or prefacing as needed. This wasa good overview of the ERIC technique (Explore, Refine, Influence, and Confirm).
Well, that wraps up my week here in Los Angeles.Special thanks to my two colleagues, Jack Arnold and Glenn Hechler, both from the Tucson Executive Briefing Center,who helped me prepare and review my presentations!