Tuesday Jun 12, 2007
www.flickr.com
|
Malhar Anaokar
SeChang Oh
Ken Pepple
Divyesh Shah
Takashi Shitamichi
Saday Tiwari
Today's Page Hits: 45
Tuesday Jun 12, 2007
I am an unabashed fan of processor design, and the past few years have been exciting times. The most recent processor introduction (the POWER6 from IBM, in case you had gone walkabout over the last fortnight) was much awaited then, by processor aficionados.
 |
This is a simplification, of course - techniques like larger caches, pre-fetching, super-scalar execution, deeper pipelines, simultaneous multi-threading, etc., are crammed into the chip to try and hide the memory latency problem. The issue is that these can be expensive to research and implement, and one rapidly runs out of newer ways of hiding a problem one can't solve (Watch this space though).
The alternate is an option pioneered by the UltraSPARC T1. Thread level parallelism combined with 8 cores (each can switch amongst 4 threads with zero latency) on a chip replace all of the techniques mentioned above, and deliver a system throughput that 2 years ago was 15x more than an equivalent chip was delivering 4 years ago. Not content with 32 threads, we will soon be launching systems based on a 64 thread successor.
Besides, even all of those techniques do not result in squeezing everything out of the increase in processor frequency. A classic example is the POWER6 itself. At 4.7GHz its frequency is 2.13x that of the POWER5+ at 2.2GHz, but performance went up only 1.4-1.45x, according to IBMs p5 570 and the POWER6 based 570 rPerf indicators. Further, at a reported 160W the 4.7GHz POWER6 chip consumes twice the wattage of the UltraSPARC T1.
What does all this have to do with iTunes? For reasons completely undecipherable, the IBM press announcement had the following statement :
the processor bandwidth of the POWER6 chip — 300 gigabytes per second — could download the entire iTunes catalog in about 60 seconds — 30 times faster than the Itanium processor in H-P's servers.
This gave me pause. A chip can do that? How interesting. To see whether this is possible, I ran a thought experiment. The most famous practioner of thought experiments was Einstein, who ultimately produced the Special Theory of Relativity as a result of one. Far be it from me to claim anything remotely as revolutionary, but what I came up with was the following diagram of the POWER6.
 |
I then tried downloading 300GB per second into that chip, ignoring the tiny problem of where I would store it while downloading the next 300GB chunk as time ticked on. There is the GX+ I/O bus, but that gives me only 20GB, assuming all of the bandwidth is for iTunes downloads. The L3 cache and Main Memory are just intermediate storage locations for the same data. Perhaps we can use those to temporarily store songs, especially because there is only 20GB coming in so far. Never mind that the processor may not be able to process 20GB of packets coming in and push them into memory in the same second. There are the inter- and intra- node buses but they only connect to other chips. Unfortunately, IBM does not publish delivered I/O throughputs, so my thought experiment comes to a grinding halt, breaking down at 20GB/s as the theoretical best one chip can do.
Perhaps we can exploit the GX+ buses on other chips to pump in the remaining 280GB? That would mean we need 15 chips to connect to each other, and then some incredible system and OS efficiency to saturate 15 sets of system buses. One could spend a lot of pre-thought experiment time cramming sections of the 18 Terabytes into all the devices on all the I/O buses. Hmm, the two POWER6 processors on our particular chip will then pack up and apply for entry to the Elysian Fields. Assuming they thus began receiving the 18TB, where would they put it? The POWER6s on the other chips will have commenced a clock-down strike, demanding to know why our chip is the privileged recipient, while they are slaves working for a song.
You might admonish me - press releases are full of sound and fury signifying nothing, you might say. They should not be read (atleast not literally) by people who have nothing better to do than conducting frivolous thought experiments. My only riposte to this is that I did manage to get a blog post out of it all.
I must confess that I got the details on the POWER6 chip from Bradley McCredie's POWER roadmap presentation - slide 7. Some of the bandwidth numbers are CPU frequency dependent, and lesser at 3.5 and 4.2GHz.
Thursday May 24, 2007
One of the top hits on Google for the phrase cpu hot plug is titled Sun CPU Hot Plug. Since it originated in the ibm.com domain, I promptly read the document, which is part of a wiki called IBM System p Expert Corner for Business Partners (I wondered how a random Googler could be a business partner, but this is the Participation Age after all).
The document has been around since at least January 2006, and is very interesting. It starts off cryptically (I am reading Brian Greene's The Fabric of the Cosmos and was thinking that the sun loses mass) but the title anchored me to the context.
"When Sun loses a processor (not deallocates a processor), the application goes down and has to be recovered. The domain that contains the failed Uniboard can restart with one less CPU. i.e. An unplanned outage that requires rebooting Solaris and recovering the application."If one were to replace Sun with IBM, Uniboard with MCM, and Solaris with AIX, the statement would still be true, but be that as it may. It then goes on to say
"Key to the discussion on hot-plug CPUs (with Sun bigots) is to give credit to ADDING additional resources to a system without having to take down the entire system (for now, we accomplish this via CUoD)."I am a Sun bigot, I guess, and in this hypothetical discussion will conditionally accept credit for adding resources to a running system. The document does not mention that we can safely add resources physically while the system is running (since the mid 1990s). Neither is our ability to physically remove resources acknowledged, but it is casually mentioned in the next paragraph -
"... Assuming the domain restarted with one less processor (described above)... once you have removed the old and added the new Uniboard to the system, adding the processors to the application environment could force yet another restart of the applications"
Err. My customers presumably were unaware of this : they have (for the past five years in my experience) dynamically reconfigured Uniboards on production systems without shutting down applications. Heads are going to roll! The pivotal word in that sentence is could, I suppose. However the next section decides to discard such pussy-footing ambiguity :
"These forced, multiple application and potentially domain outages to add or remove a resource to a domain (or a server with one domain) makes their Containers argument much less viable."
It does characterize the domain outages as potential. These application/domain outages must be why IBM has not had the equivalent of Dynamic Reconfiguration on AIX systems this far. Never mind that Solaris provided DR since 1995, and customers have been using it effectively ever since, in production and with applications on-line.
On the other hand, in the light of unviable Container technology, why is IBM announcing Workload Partitions in the AIX 6 Beta? Multiple App restarts, Domain outages, adding and removing resources .. what is AIX coming to?
Naughty, naughty.
Speaking of AIX 6 : Open Beta Program, Workload Partitions, Live Application Mobility, Role Based Access Control, Trusted AIX, Secure By Default Install Option, Dynamic Tracing ... notice something? Next thing you know, they will release AIX 6 code under an OSI approved license.
Tags : solaris aix partitions
Tuesday Feb 27, 2007
Ken Milberg compares virtualization offerings from IBM, HP and Sun in his article for IBM Systems Magazine. I was beginning to get into details in this post, when I realized Jim Laurent has already done so here. I restrict myself therefore to the Summary in Ken's article.
Capped and uncapped partitions. Allowing users to take advantage of unused clock cycles via a shared processor pool is an innovation that no one else has. HP requires a workload manager system similar to PLM, while Sun has nothing.
According to the Advanced POWER Virtualization on IBM System p5 Redbook draft, capped mode will limit processing units given to the partition to the entitlement guaranteed by the system, even when resources are available in the shared processing pool). In the uncapped mode, the processing capacity given to a partition may exceed the guaranteed processing capacity when resources are available in the shared processing pool. One must specify the uncapped weight of that partition, so that if multiple uncapped partitions demand idle resources, the allocation will be in proportion to their weights.
I am guessing Ken agrees that the uncapped mode is trivial, and the default behaviour in an approach to allocating resources to multiple applications. The Fair Share Scheduler acts as the weighting mechanism in Solaris Containers. As for Capped Mode, I am unsure how Ken believes this is different from the concept of Capped Containers, a mode where a Container cannot obtain more than the processing capacity guaranteed to it.
SMT - Only IBM has it
Since this is like saying CMT - Only Sun has it, and Solaris Containers exploit it, I cannot understand the relevance. I went back to part 1 of Ken's article to see if he describes how SMT delivers value unique to Advanced POWER Virtualization, but could not find detail.
Dedicated or Shared I/O on a virtual partition - Only IBM has it
In part 1 of his article, Ken elaborates :
A physical adapter assigned to a partition can be shared by one or more other partitions.. The VIO Server is designed to reduce costs by eliminating the need for dedicated network adapters, disk adapters and disk drives.. Devices can be a mixture of dedicated devices assigned to partitions for maximum performance, or used in the Virtual I/O Hosting Partition to be shared by multiple partitions to provide higher efficiency of resources and adapters.
Do we have something called a VIO Server in Solaris Containers? No, we do not. Do we have the ability to dedicate or share I/O resources? Yes, we do.
IBM has only one virtualization strategy (APV)
I agree, and in fact, I believe this is a limitation. What if customers want the ability to electrically isolate application and operating system environments? Advanced POWER Virtualization does not deliver it, and customers will have to buy multiple pSeries systems to deploy those environments. What if customers want the ability to isolate application execution environments, but without the overhead of hypervisors and the like? IBM's virtualization product does not satisfy this need either.
Ken, on the other hand, believes that Dynamic System Domains, Solaris Containers, Logical Domains et al are an attempt to confuse customers. I think this opinion "misunderestimates" the ability of customers to grasp the concepts and choose one that suits a specific requirement.
One hardware platform for AIX and Linux partitions (POWER5)
I am not sure why one would buy relatively expensive p5 systems to run Linux, especially if the applications available are limited, and undergo a porting/certification cycle separate from the x86/x64 & Linux porting cycle, making for extra work for a software vendor. Without entering that debate, Solaris Containers for Linux Applications allow customers to, if they need it, run Linux Applications (initially those certified for RHAS3 and CentOS) unmodified in Solaris environments. The technology has already been released through Solaris Express, and will be coming soon to a Solaris 10 Update near you.
Ken's Solaris buddies should be guiding him to locations like the Solaris Containers Learning Centre, rather than telling him Windows runs on the SF T2000.
The article does acknowledge some of the strengths of Solaris Containers over Micropartitions/lPARs- it runs across both SPARC and x86/x64 platforms, and it imposes no overheads.
I think providing virtualization capabilities that span Resource Management, Isolated Application Environments, Software Partitions and Hardware Partitions is a good thing. Likewise, I think providing as many of these capabilities as is possible across multiple platforms is equally brilliant. With enthusiastic community backing, endorsement by Intel, IBM, and HP, and strong adoption, our R&D engineers will continue to enhance features across the spectrum.
Update : Jeff Savit comes to the party.
Tags : virtualization; solaris containers; solaris
Sunday Feb 25, 2007
Nostalgia isn't what it used to be. However, if you are curious about Sun's origins, the ten minute video alongside has Scott McNealy talking about how the company began. His Big Mac moment, Bill Joy coming on board (and later, wanting to use the acronym BUN rather than SUN - was this association with bread where the pizza box form factor idea came from?), building a vi driven ERP system, and the ComputerVision deal feature prominently.
Want more? The Sun Founders' Panel event, also from last year, was part of the Computer History Museum Odysseys in Technology series. Andy Bechtolsheim, Bill Joy, Vinod Khosla and Scott McNealy spent a couple of hours with John "the Network is the Computer" Gage, talking about the company they founded, and computing in general. The video, amongst other McNealy podcasts, presentations and interviews, is here.
Tags : sun
Trail Blazers
What is common to Ibn Battuta, Alain Prost and Sun Microsystems? Their birthday - 24th February.
Ibn Battuta covered over 75000 miles in his epic journeys - spanning a geography now spread across 44 countries. His account of the places he visited are today considered the most authoritative 14th century overview of the region stretching from Spain to China.
Alain Prost must have driven nearly as much distance on Grand Prix weekends in his 202 Formula One races, but he is best known for his jousting with Ayrton Senna through a decade of fascinating racing.
Sun has covered a lot of ground over the last 25 years, with arguably the broadest Open Source underpinnings in the corporate world, and some of the best minds in the industry giving shape to a vision that seems speculative for when it was first framed, but now sounds like the defining statement about the past two decades. As the lines between the sources of data, the means by which it is processed into information and the destinations to which it is disseminated, get blurred (dare I say The Producer is the Consumer), the vision statement becomes a truism.
A lot of the buzz around Sun has been centred on the technologies and standards that it developed and/or evangelized - TCP/IP, NFS, RISC, UNIX, Java, NetBeans, Project Liberty, OpenOffice, GlassFish to name a few. A bit has also been sparked by the battles that we have picked to fight : rivalries that everyone in the technology industry and from elsewhere have followed and, let us admit it, enjoyed.
I watch a Grand Prix race not as much to see 20 odd cars speed metronomically around the circuit and one emerge fastest, as much as it is to witness absorbing combat between drivers; to delight in strategy, tactics, team work, brains, brawn (I am not speaking of Ross) and machine coming together to execute a poetic overtaking manoeuvre at a chicane. Sometimes it works and other times it does not, but the attempt is often absorbing enough. And there is always the next lap, or the next race, or the next season for one more go.
As you might have guessed, I can't wait for 18th March. Michael Schumacher won't be behind a wheel, but he will contribute to the strategy, tactics, team work and brains part of Ferrari.
Meanwhile, it is 24th February. Happy Birthday, Sun, from a planet on which the sun will never set (or atleast, not for a few billion years).
Tags : sun
Tuesday Feb 13, 2007
The first of a series of Sun Blueprints on the Tokyo Institute of Technology Supercomputer Grid has been published. Written by Nobu Hashizume-san, this article focuses on the institute's requirements, and architecture of the grid. It also discusses the optimizations used to achieve the LINPACK results that propelled the grid to ninth in the Top500 Supercomputer sites in the world.
The Titech supercomputer comprises around 655 SF X4600s and 42 SF X4500s, along with Voltaire Grid Director Infiniband switches and ClearSpeed Advance Floating Point PCI-X accelerators. Combined with N1 Grid Engine, N1 System Manager and other software, the grid delivered 47.38 TFlops in November 2006.
The Grid was named Tsubame, Japanese for Swallow (as in the bird). I suspect it is named so after the Tsubame trains that plied between Tokyo and Osaka in the 1950s, and were the fastest at the time.
Tags : top500; supercomputer; tsubame
Wednesday Feb 07, 2007
As I begin this post, Greg Papadopoulos is concluding his SAS 2007 address. It is always interesting to watch GregP paint a picture of the future, with broad brush strokes that somehow introduce subtle variations here and there. Parts of his presentation re-iterated ideas in some of his blog posts - particularly The Ecology of Computing.
The Redshift is a property of light equivalent to the Doppler Effect property of sound, and occurs when a light source moves away from the observer : the light emanating from the source shifts toward the Red part of the VIBGYOR spectrum. It became the basis for Hubble's law, and is often cited as evidence for an expanding universe.
As it applies to systems, Greg used the term as a leading observation of the expanding market. He began by establishing magnitudes of increase in compute throughput capabilities over time and contrasted it with the relatively sedate world economic growth in the same period. The question then is what or who is fueling the seemingly disproportionate demand for computing? One answer is the proliferation of client devices that require backend systems to obtain information from, communicate with, and send back telemetry to. Another angle to it is High Performance Computing - organizations weren't doing some processing that they would have liked to, because either there wasn't enough compute capacity to do it, or it was economically/ecologically unviable to harness that capacity. Over a period of time, Moore's Law makes it feasible, and HPC customers commence running those tasks.
The third category, one that Greg called *-Prises or Star-Prises, are organizations like Google or Sun's own Network.com - aggregators of function or capacity that individuals or enterprises used to own earlier. A Salesforce.com, for instance, will obviate the need for a start-up to build it's own CRM infrastructure. Given the number of start-ups Palo Alto cafes spawn, and the number of mature organizations that think CRM is not a core competency, a provider of CRM as a service could grow at a clip that Moore's Law can no longer keep pace with. Voilà, *-Prise.
All three sectors view IT as a competitive weapon. There will be sectors that consider IT as a cost centre, and Sun continues to innovate in areas like consolidation, virtualization and power-efficient computing to address those.
Chip Multi-Threading and Project Blackbox (by the way, is Blackbox coming to a city near you?) are two responses by Sun to the Redshift. Combined with an Operating System that eats threads for breakfast, they form a perfect platform for those who are under-served by Moore's Law.
Greg also described the Neptune and Crossbow projects - designed to cope with the exploding network bandwidth demands. To continue borrowing from astronomy, I see the Redshift as an Event horizon, which is a sort of cosmological Tipping Point. It would be incautious to persist with the metaphor though, because it leads to the point of no return!
Thursday Jan 18, 2007
The Bulletin of the Atomic Scientists moved the minute hand of the Doomsday Clock two minutes closer to midnight, symbolizing an increased threat to humankind's survival.
Significantly, the magazine of nuclear physicists has begun tracking climate change, in addition to nuclear weapons, explaining that
Climate change also presents a dire challenge to humanity. Damage to ecosystems is already taking place; flooding, destructive storms, increased drought, and polar ice melt are causing loss of life and property.
The thinking resonates with another story that described the "creation" of a new island off Greenland, due to Arctic melting. Closer home, the average temperatures in India cities have perceptibly increased. Bangalore's summers are no longer as pleasant as they used to be hardly a decade ago.
I am glad that products like the UltraSPARC T1 and Niagara2, and concepts like Blackbox contribute to ecologically responsible computing.
Tags : environment; coolthreads
Monday Dec 04, 2006
In a fascinating article, Clive Thompson of the New York Times describes how the US Intelligence Agencies are evaluating "Web 2.0 technologies" to come up with timely, authoritative analyses of events. I like the point two of the people involved made : In a world awash in information, the meaning of intelligence is shifting. Like all enterprises dealing with today's rapidly changing world, the Intelligence Community is faced with the challenges of converting data into useful information that can be used to obtain insight.
The article describes efforts to use Blogs and Wikis, and also a comic (sadly) situation resulting from the use of inappropriate filters. Apparently, five years ago, classified information wasn't being eliminated from some data, as it should have been. It was discovered that the filter used was for the classification "secret", and so, any data marked SECRET or S E C R E T would escape the filter.
Thursday Nov 30, 2006
I will be speaking to a new recruit soon as part of his orientation program. The program is informal, but I believe it sets the tone for the person's career in Sun, and is the filter through which he will perceive much of what we do and communicate. First impressions, like they say.
As Chief Technologist for India, I expect to spend time on our products and solutions. However, Sun is much more than the sum of our technologies. Don’t get me wrong : Solaris, Java and SPARC, for example, are near iconic in their functional areas, I think. However I should be starting with what their design principles collectively embody.
I post my thoughts here, because far more accomplished minds than mine have convinced me that Innovation Happens Elsewhere, and that the Many are Smarter than the Few. The feedback I receive should be interesting.
Several taglines describe what Sun does, but I am looking for why we do what we do. Fundamentally, our objective is to enable every man, woman and child to become more aware, more knowledgeable. In today's world, this translates into eliminating the digital divide. The personal computer, and its cross bred offspring the cell phone, have surpassed any means of access and display we thought of. The Internet has overwhelmed every distribution mechanism we have invented. The Web has rendered (pardon the pun) any other presentation layer primitive in scope. Facilitate the growth and spread of the network, and universal (well, planetary - though some stuff I come across on the Web can only be conceived by alien intelligence) access to knowlede follows.
This is particularly relevant in my nation. Slums rubbing metaphorical shoulders with posh residences is only the first of several, sadly cliched, expressions describing the divisions that threaten our collective progress. Many organizations and individuals aim to achieve an inclusive society here, and Sun India collaborates with some of them, but that is a subject for another post.
That is our Cause then, but how will we succeed in this mission to provide information access and enable people to participate in communities that foster knowledge and exchange ideas? Sun recruits bright minds to solve digital information infrastructure problems together. We outfit the world with the means to store, process, access and deliver digitized data and information. Increasingly, this infrastructure is aggregated by large enterprises, service providers and governments - economies of scale kicking in, I guess.
We focus on Information Technology. To mangle an old saying, the Sun never sets on the Information Empire, and the idea is to convert the oligarchy into a democracy. That is stretching the symbolism a bit too far, do you think? IT, by the way, is an interesting term - combining both the end (Information) and the means (Technology). The Information in IT is sometimes overshadowed by the Technology, but we try not to lose sight of the purpose for which we develop the technology.
That in a, rather large, nut-shell is how I have framed the opening gambit in a friendly game of chess with our new Technical Architect. I look forward to his opinion too, especially as he spurned an offer from a competitor to join us.
This post is part of the Output I refer to in the title of this blog (the tech in cheek reference to the Samuel Beckett play is unintended). It is the Input half I am most excited about; the tedium that Vladimir and Estragon represent will hopefully never show up here.
Tags : sun