Thursday, 17 August 2017

The systems-area obsession with peak performance

In systems, there has always been a completely understandable focus on peak performance.  For me personally, Butler Lampson's early papers on "Hints for Operating System Design" (which basically argued that we need to unclutter the critical path), the famous End to End paper by Saltzer, Reed, and Clark (which argued for taking unnecessary functionality out of the network), and the Birrell and Nelson paper on the performance of Firefly RPC (which argued for taking unnecessary mechanism out of the remote procedure call path) were awe-inspiring classics: papers you reread decades later, and that still amaze.


In fact for people who get pleasure from programming, there is a natural tendency to build systems and evaluate them, and obviously any such task centers on a peak performance perspective.  Optimizing peak performance is fun, and honestly, can be addicting: there is such a direct feedback when you succeed.  It is very rare to see a first version of a system that can't be sped up by a factor of 10 or more even just by doing basic optimizations, and in some cases, we end up with speedups of 100 or 1000-fold, or even more.  What a thrill!
Yet there is a sense in which our love for speed worries me: I wonder if this classical way of thinking about systems might be fading as a pure form of innovation in the eyes of the field as a whole.  The core issue is that existing systems (here I mean the mainstream workhorses: the operating system, the network layer, perhaps the compiler) all work pretty well.  Of course, speeding them up is a worthy endeavor, but it may no longer matter enough to be a justifiable goal in its own right.  Research on performance is just not a compelling story, if you focus on this particular layer.
Why should this matter?  In fact, it might not matter at all, were the systems community itself aligned with these larger external forces that shape the way we are perceived by other communities and by computer science as a discipline.  But right now, I suspect, there is a substantial disconnect: people like me are addicted to speed (hmm... that doesn't sound quite right), while people who hang out at conferences like NPS and KDD don't really spend much time worrying about the performance of the systems components my crowd focuses upon, like the latest version of Linux running on the latest multicore hardware platform.
As I write this blog entry, this same dynamic is evident even within my own research group.  For us, it plays out as a tension between telling the Derecho story as a story about a new concept ("smart memory") and telling it as a story about raw speed ("fastest Paxos and Atomic Multicast, ever!"). 
It seems to me that the broader field tends to value  impact more than what might be called "narrow" metrics, such as the speed of the Linux I/O path.  Invent a new way of doing things, and you can change the world in interesting ways.  So the puzzle that emerges is this: if the systems community has started to drift relative to the broader computer science community as a whole, don't we then run some risk of becoming marginalized, by virtue of over-emphasizing aspects that the broader computer science community views as unimportant, while actually rejecting innovations that the broader community might be thrilled to hear about?

Take Spark, a recent home run story from Berkeley.  If you think back, the first research papers on Spark told a story (primarily) about huge speedups for MapReduce/Hadoop, which  obtained by smarter in-memory caching of files (they call them RDD: Resilient Distributed Data objects) and smarter task scheduling, so that computations would tend to exhibit affinity relative to the cached content.  Years later, it seems clear that the more significant aspect of Spark -- the more impactful innovation --  was that it created a longer term computing "model" in which data loaded into the Databricks system (the new name for Spark) lives in memory, is transformed through various stages of computation, and where the end-user has a powerful new experience of data mining with vastly better performance because these RDDs remain resident in memory as long as there is enough space, if they continue to be used now and then.  Systems people may think of this as a story of performance... but NIPS and KDD people perceive this as a paradigm shift.  As a systems person, it seems to me that our community in fact accepted the right papers, but for the wrong reason, and in fact that the early advising of the Spark inventors (by their faculty mentors) may even have misunderstood the real contribution and steered them down the less vital path.  Of course, the enthusiasm for Spark quickly reset the focus, and today, the Databricks company that offers commercial support for the Spark platform focuses on high-productivity data mining with blazing performance, rather than portraying the story as "speeding up your Hadoop jobs."


It isn't a realistic thing to wish for, but I'll wish for it anyhow: as a field, it seems to me that we need to try and pivot, and to embrace change in terms of what styles of computing really matter.  The world has changed drastically in the past decade or two: what matters most, right now, is machine learning.  This is partly because systems work pretty well.  Disruption comes from big reaches, not small 10x optimizations to things that already worked fairly well.


I don't know anything more about the future than anyone else.  My focus, in Derecho, is on "smart memory," but will this ever become a recognized field of research, one that other people would work on?  Does the world really need smart memory services for IoT applications?  I hope so, but of course that question will be answered by other people, not by me.    And so one can easily understand why my students love the raw speed story: For them, fast replication is a more well-defined systems topic, with an obvious and specific role in existing systems.  People use replication solutions.  So it makes sense for them to gravitate towards speed records.


Indeed, for them, viewing machine learning as the real goal, and performance as just one dimension, makes systems research feel secondary to machine learning research.  Nobody wants to feel like the plumber or the electrician: we all want to build the house itself.  Yet perhaps this is the new reality for systems researchers.
Will such a pivot be feasible?  Perhaps not: the systems addiction to speed runs deep. But at the same time, when I visit colleagues in industry, I find them deeply embedded into groups that are doing important practical tasks that often center on a machine learning objective.  So it seems to me that if we don't evolve in this necessary way, we'll simply fade in importance to the broader field.  We just have to try, even if we might not succeed.

Thursday, 10 August 2017

Zero-copy computing

Last night, one of my group members tossed together a simple experiment on Derecho as the first step towards a much fancier experiment he needs to run.  To his surprise, the performance was a fifth of what we've been seeing in our experiments that will go into the ACM TOCS submission we plan to send out any day now.


What could cause a 5x slowdown?


As it turns out, the issue is easy to understand and points to a deeper and rather interesting puzzle.  RDMA is blazingly fast, as you know if you've read my earlier postings on the topic.  Basically, transfers occurs entirely in hardware: the NIC on machine A talks to the memory unit on machine A and grabs chunks of data, which zip straight over the wire to the NIC on machine B, which stores the data directly into the memory unit on B.  This gives a rate of data transfer that can be much higher than any single-core memcpy operation could approach: even if there is a hardware instruction for copying blocks of memory, that instruction still will operate by a series of fetch and store operations and will need to talk to the memory unit twice for each cache-line-sized chunk of data.

So RDMA tends to run 2x faster than copying.  If you look at an end-to-end pipeline, you'll generally see that RDMA delays are sharply higher than the latency of interacting with local DRAM, but the data transfer speed for a big transfer can maintain this 2x benefit, from data on A all the way to the receive buffer on B.


This is what my student ran into last night.  In his case, by using Derecho to send long null-terminated strings he benefitted because strings are easy to check for correct content ("Hello world, this is update 227!") but very costly to create and transmit.  Derecho was running 5x too slowly because it spent all its time waiting for his expensive string creation code to run, and for Derecho's orderedsend to marshall the objects before sending.  Yet our instinct tells us that those should be viewed as fast operations.  Well, instinct has ceased to be correct: RDMA is a new world.


In fact you've probably thought about the following question.  You are driving on highway 101 in a fancy Tesla sports car with one of those insane speed buttons.  Elon Musk has really pushed the limits and the car can reach 2/3rs of the speed of light.  Yet your commute from Menlo Park to South San Jose takes exactly as long as it took back when you were driving your old Subaru Imprezza!  The key insight isn't a very deep one: even with a supercar, the "barriers" on the highway will still limit you to roughly the same total commute time, if those barriers are frequent enough. 

In modern operating systems and languages, these kinds of barriers are pervasive.


Today's most widely used standard systems maintain data in various typed data structures: strings, classes defined by developers, etc.  Even an object as simple as a string may require byte by byte copying, just to find the null terminating character, and by itself, this will be a further 2.5x slower than RDMA.  So send a string, and your end-to-end numbers might be 5x slower than the best we get from Derecho with byte arrays of known size.  Send a class that needs complex marshaling and the costs go even higher (if the fields can be copied directly from memory, scatter-gather is an option, but otherwise we would often need to first copy the data into a send buffer, then send it, then free the buffer: a sequence that could push us even beyond that 5x slowdown).


What you can see here is that at speeds of 100Gbps or higher, copying is a devastingly slow operation!  Yet in fact, modern operating systems copy like crazy:
  • They copy data from user space into kernel space prior to doing I/O, and back later.
  • Modern languages are very relaxed about creating clones of objects.
  • Other than C++ 14, every method call copies arguments onto the stack, item by item.
  • Garbage collectors copy and compact all over the place.
This suggests that it may be time for someone to create a zero-copy operating system, and perhaps also to start looking carefully at what it would take to do zero-copy versions of modern programming language frameworks like Python, Go, C# and so forth (C++ is pretty good in this way).  Otherwise, as RDMA pushes towards 400 Gbps (200Gbps is already becoming fairly common), and 1Tbps within a decade, we'll find that our RDMA path seems nearly instant... but that the applications just can't benefit!

Monday, 31 July 2017

Why is it so hard to mask failures?

When we talk about fault tolerant distributed computing, using the state machine replication approach, it may seem obvious that a system of this kind should be capable of completely masking failures.  In fact, however, this is not the case.  Our ability to hide failures is really very limited.

When developers use state machine replication techniques (SMR), the usual approach is to replace components of systems or distributed services with groups of N members, and then use some sort of library that delivers the same inputs to each, in the same order.  If the replicated component is deterministic, and if care is taken to initialize each component in a manner properly synchronized with respect to its peers, this is enough to guarantee that the copies will remain synchronized.  Thus, we have an N-replica group that seemingly will tolerate N-1 faults.

Unfortunately, theory is one thing and reality can be quite a different matter.  When people first began to experiment with SMR in the 1990's, developers quickly noticed that because software bugs are a major cause of failure, perfect replication will replicate many kinds of faults!  Over time, a more nuanced approach emerged, in which the various replicas are proactively shut down and restarted in an uncoordinated way, so that on average there would still be N copies, but at any instant in time there might be N-1 copies, with one copy shutting down or rejoining.  The trick is to transfer the current state of the group to the recovering member, and is solved using the virtual synchrony model, in which group membership advances through a series of epochs, reported via view upcall notifications, with state transfers performed during epoch transitions.

The benefit of this sort of staggered restart is to overcome so-called Heisenbugs.  The term refers to bugs that are hard to pin down: they could cause non-deterministic behavior (in which case the replicas might diverge), or bugs that seem to shift around when the developer tries to isolate them.
A common form of Heisenbug involves situations where a thread damages a data structure, but the damage won't be noticed until much later, at which point any of a number of other threads could try to access the structure and crash.  Thus the failure, when it occurs, is associated with logic remote from the true bug, and may jump about depending on scheduling order.  If the developer realizes that the root cause is the earlier damage to the data structure, it generally isn't too hard to fix the problem.  But if the developer is tricked into thinking the bug manifested in the code that triggered the crash, any attempts to modify that logic will probably just make things worse! 

The reason that staggered restart overcomes Heisenbugs is that a restarting program will load its initial state from some form of checkpoint, hence we end up with N copies, each using different operations to reach the same coordinated state as the other N-1.  If the data-structure corruption problem isn't a common thing, this joining process is unlikely to have corrupted the same data structure as did the others.  With proactive restart, all N copies may be in equivalent yet rather different states.  We can take this form of diversity even further by load-balancing read-requests across our N copies: each will see different read operations and this will be a further source of execution diversity, without mutating states in ways that can cause the N replicas to diverse.
With such steps, it isn't hard to build an ultra-resilient SMR service, that can remain alive even through extremely disruptive failure episodes. But can such a service "mask" failures?

The answer is yes and no.

On the "yes" side we find work by Robert Surton at Cornell, who created a clever little TCP fail-over solution called TCP-R.  Using this protocol, a TCP connection can seamlessly roll from one machine (the failed server) to another.  The cleverness arises because of the way that TCP itself handles acknowledgements: in Surton's approach, a service member accepts a TCP connection, reads the request (this assumes that the request size is smaller than the TCP window size, in bytes), replicates the received request using an SMR multicast, and only then allows TCP to acknowledge the bytes  comprising the request. 

If a failure disrupts the sequence, TCP-R allows a backup to take control over the TCP session and to reread the same bytes from the window.   Thus the service is guaranteed to read the request at least once.  A de-duplication step ensures that a request that happens to be read twice won't cause multiple state updates.

Replies back to the end user are handled in a similar way.   The service member about to send the reply first notifies the other members, using a multicast, and only then sends the reply.  If a failure occurs, one of the other members claims the TCP endpoint and finishes the interrupted send.
With TCP-R the end-user's experience is of a fully masked failure: the application sends its request, and the service definitely gets the request (unless all N members crash simultaneously, which will break the TCP session).

Lacking TCP-R, the situation is quite a bit more complex.  In effect, the end-user would need to send the request, but also be prepared to re-send it if the endpoint fails without responding.  For read-only requests, the service can just perform the request multiple times if it shows up multiple times, but updates are more complex.  For these, the service would need logic to deduplicate requests: if the same request shows up twice, it should resend the original reply and not mutate the service state by performing the identical operation a second time.  TCP-R masks the service failure from the perspective of the client, although the service itself still needs this form of deduplication logic.

On the "no" side of the coin, we need to consider the much broader range of situations that can arise in systems that use SMR for fault-tolerance.  In particular, suppose that one SMR-replicated service somehow interacts with a second SMR-replicated service.  Should all N members of the first replica group repeat the same request to the M members of the second group?   Doing so is clearly the most general solution, but runs into the difficulty that bytes will be transferred N times to each member: a high cost if we are simply trying to mask a rare event!

Eric Cooper studied this question in his PhD thesis on a system called Circus, at Berkeley in the 1990's.  Basically, he explored the range of options from sending one request from group A to group B, but reissuing the request if the sender in A failed or the receiver in B, all the way to the full N x M approach in which every member of A multicasts every request to every member of B, and the members of B thus receive N copies and must discard N-1 of them in the usual case.  (TCP-R can be understood as an instance of the first approach, but with the client-side logic hidden under TCP itself, so that only the server has to be aware of the risk of redundancy, and so that it only arises when a genuine failure occurs.)

Cooper pointed out that even with high costs for redundancy, the N x M approach can often outperform any scheme that waits to sense a failure before retrying.  His insight was that because detecting a failure can be slow (often 30s or more), proactively sending multiple copies of each request will generally produce extra load on the receivers, but with the benefit of ensuring a snappy response because at least some receiver will act promptly and send the desired reply with minimal delay.

Cooper's solution, Circus, is best understood as a design pattern: a methodology that the application developer is expected to implement.  It involves a multicast from group A to group B, code in group B to remember recent responses to requests from A, and logic to de-duplicate the request stream, so that whether B receives a request 1 time or N times, it behaves identically and responds to A in the same manner.

In Derecho, we don't currently offer any special help for this heavily redundant approach, but all the needed functionality is available and the design pattern shouldn't be hard to instantiate.  But in fact, many Derecho users are more likely to use a non-fault-tolerant approach when building a data processing pipeline.  More specifically, while Derecho users would often use replication to make the processing elements of the pipeline fault-tolerant, they might decide not to pay the overhead of making the handoff of requests, stage by stage in the pipeline, ultra-reliable.

The reason for this compromise is  that in the IoT settings where a "smart memory service" might be used, most sensors are rather redundant, sending photo after photo of the same car on the highway, or location after location for the cat in the apartment.  The service receives this heavily duplicative input and will actually start by sorting out the good content and discarding the replicated data.  Thus we suspect that most Derecho users will be more concerned with ensuring that the service itself is highly available, and less concerned with ensuring that every single message sent to the service is processed. 

Indeed, in most IoT settings, freshness of data is more important that perfect processing of each and every data point.  Thus, if some camera creates photo X of a vehicle on the highway, and then photo Y, and X is somehow lost because of a failure that occurs exactly as it is being sent, it often would make more sense to not fuss about X and just focus on processing request Y instead.

Microsoft has a system, Cosmos, in which a pipeline of processing is done on images and videos.  It manages without fault-tolerant handoff between stages because failures are rare and, if some object is missing, there is always a simple recipe to create it again from scratch.  Facebook apparently does this too.  Both systems need to be extra careful with the original copies of photos and videos, but computed artifacts can always be regenerated.  Thus, perfect fault tolerance isn't really needed!

Of course, one can easily imagine systems in which each piece of data sent to the service is individually of vital importance, and for those, an open question remains: is it really necessary to hand-code Cooper's Circus design pattern?  Or could there be a really nice way to package the Circus concept, for example by using a higher level language to describe services and compiling them down to SMR replicas that talk to one-another redundantly? 

I view this as an open research topic for Derecho, and one we may actually tackle in coming years.  Until then, Derecho can certainly support a high quality of adaptation after crashes, but won't seamlessly hide crashes from the developer who works with the technology.  But on the other hand, neither does any other technology of which I'm aware!


Saturday, 15 July 2017

How far could a secure Internet get us?

There is a standard refrain among those who teach and do research on security: it holds that the fundamental flaw in security originates with an early decision by the community that created the Internet.  In that early design, Internet routing occurs by a form of peering that can often occur anonymously.  Endpoint systems are not authenticated, and packets generally aren't encrypted or authenticated either.  As a consequence, packets are routed permissively even if they contain what are almost certainly false origin addresses, or have content that seems dubious in obvious ways (like a TCP connection packet that makes no sense on this particular link).

This story then holds that the original sin of the Internet design community has given us layer upon layer of infrastructure that cannot be secured because it resides on a flawed and easily compromised foundation.

As it turns out, there is little doubt that we actually can build an Internet that corrects that deep flaw, and that can be relatively compatible with the existing Internet.  Enterprises like military systems and security-conscious companies do it all the time.   Let me sketch out how this is done (it isn't rocket science, and you won't find many surprises).  But then let's ask if it would really be a game-changer for the security of higher level applications.

To appreciate the core challenge, it helps to start by understanding that everything you think you know about the network is really just an abstraction layered over something more basic.  For example, even my advanced students tend to think of Internet addresses as identifiers, as if my computer's true name were 128.84.16.001 and your computer's true name was 167.88.194.023.  In fact, these addresses are better understood as routing data: useful for deciding what the next hop should be as a packet progresses hop-by-hop through the network, but lacking any deep connection to identity. 

In fact, a single computer might have many network addresses (one for each of its network interfaces).  It might host multiple virtual machines, and those could each have virtual network addresses of their own: with virtualization, any computer can function as a kind of Internet and host large numbers of computers that run within it, complete with virtual networks, virtual name spaces, services of various kinds, you name it.  A single computer might move about, so that over short periods of times it takes on a sequence of addresses: the old addresses can cease to work, and traffic needs to be redirected to the new ones.  Addresses can be mapped by network address translators.

Once we free ourselves from this false equivalence of network address to identity, you need to ask what the real root of identity should be, in a modern system.  Here, the proper focus is on hardware security tokens combined with human users who authenticate themselves in the usual ways.  The hardware could be a built-in component of the computer, or it could be some form of plug-in.  But the key point is that when we associate authentication with these unforgeable hardware elements, used in tandem, we arrive at a much stronger notion of endpoint identity.  The deep roots of that form of identity reside in key pairs: the identity defines some form of private key, with which information can be authenticated by public components that only have access to the corresponding public key.

This then is our bootstrap opportunity: we live in a vast world of devices: computers, routers, IoT components like video cameras and smart televisions and smart cars, and each element can be understood as either being anonymous (if it lacks the ability to authenticate itself), or capable of proving that it has "authorized access" to some private key.  With that proof, we can consult a vendor-published registry and from that registry, can learn about this endpoint device.  A device can also be registered in secondary registries: when I bring a new router into my smart home, I could register my router as "Ken's router, at his summer cottage on Cayuga Lake".  And now there would be a trustworthy chain of reasoning that would let you convince yourself that certain messages were indeed sent by, or countersigned by, my router.

Sounds familiar?  It should, if you've ever brought your laptop from home to work.  Some companies won't allow you to connect the laptop at all (fearing viruses that your machine might carry), but those that do usually require precisely this sort of authentication and registration.

Given strong authentication, a second opportunity arises whenever we focus on an island of infrastructure that has a coherent owner.  Here on Cayuga Lake, my network provider is part of a broader system owned and controlled by a company that controls a fairly large regional network.  This ISP is paid for its services, and at least in theory, has complete control of every device that can connect directly to it, every router and switch it operates internally, and every network link used to glue them all together.  One can understand the ISP as a kind of military hierarchy: at the core we find a "master and commander" who has absolute control and is the sole source of permissions.  Below the master and commander are a collection of subordinates, to whom restrictive roles have been granted, and way down at the bottom are the common seaman who might permit packets originating in my router to sail the seas of this regional network -- or could block them.

As it happens, today's ISPs are somewhat relaxed about authentication, so one opportunity to secure the network would start by simply enforcing authentication when devices are connected to an ISP.  Today when I buy a new iPhone, I can simply start to use it in my home.  If I was trying to build a much more secure network, at a minimum my router might pop up a screen requiring me to fill in details: who owns this iPhone (and it might want proof: type in your password, your RSA code number, and hold your thumb on the fingerprint reader...)  Perhaps, that regional ISP would do the same and require a second level of device registration.  Military and other nationally critical infrastructure networks work precisely in this way: if you were to take a laptop into a major power plant and connect it to the Ethernet jack in the waiting room while waiting for your appointment with human resources, either it won't connect, or it will give you some form of very limited guest connectivity. 

Think back to the last time you took your laptop to work.  I bet that something along the lines just described happened then, too.  Any organization that takes security at all seriously is careful to track the devices connected to it, and to limit their networking "power".  A military system won't allow you to connect your own machine, at all.  But if you work at Cornell, like I do, you might be able to get permission -- except that your machine will end up registered for use in a specific context, such as from my office in the Computer Science building.  If I were to carry it across the street and connect it to a wall jack in the ECE department, I would be back to square zero.

With enclaves that use authentication, one can take further steps.  Without too much added cost, packets can be cryptographically signed or fully encrypted at the endpoint, yielding various forms of virtual private networks: subnetworks within which communication is safe, but with strong protection against traffic from the outside leaking in, or against intruders managing to read or tamper with data.  Such systems can also filter or block traffic that might try to escape the secure perimeter.

I worked at Microsoft for a few months in 2016, and they adopted this approach.  I could only connect to their system via a secured VPN, and the security perimeter it enforced when I worked from home was very carefully controlled and monitored.  I could continue projects from work while at home, but I could never have wandered the virtual Microsoft network with impunity from outside the office.  In my office, I had somewhat more relaxed access to internal software tools and projects.

This, then, gives some sense of what a secure Internet would look like.  But how secure can such a system really be?

As we push from the lowest layers of hardware and software up to higher levels of abstraction, the numbers of elements of a modern system increase exponentially.  There are dozens of operating systems and each has hundreds of variants and patch levels.  So the very first layer already has a level of diversity measureable in the thousands.  Then there are hundreds of programming tools and languages and services that can run on them, configurable in hundreds of ways, to say nothing of all the management options.  Layer by layer, we bring in a surreal degree of diversity simply by virtue of the choices made by each system designer and vendor.

In settings like military systems, or power grids, a major effort is invested to keep control over the forms of diversity that are actually present in the deployed system.   Individual users aren't permitted to install their own applications, patches are applied in a coordinated way, and monitoring tools are used to notice unexpected behavior that could point to an intrusion.  In contrast, networks used in other settings need to deal with enormous levels of diversity and individual customization.  Like it or not, the network provider simply won't know what the network is being used to do.

It seems to me that this diversity accounts for the insecurity of modern systems, to a far greater degree than the "original sin" of endpoint anonymity and unauthenticated peering.  While the insecurity of the network certainly makes it easier for attackers to mount denial of service attacks or to route spam emails with falsified origin data, those are just the a small aspect of the larger problem.  And that larger problem centers on the exceptionally large attack surface that modern systems offer: hundreds of millions if not billions of lines of code, riddled with bugs, and configured in every imaginable way (just yesterday I was horrified to discover that my home router has an administrative login and password, both set to the defaults.  Now I'm wondering about my television, and my internet box...). 

Fixing the original sin of the Internet genuinely would help in some ways, especially as we move to deploy an Internet of Things with more and more devices playing a diversity of roles in homes, cars, highways and other settings.  We should take that step.  But it is an illusion to imagine that true software security can be achieved by hardening the network itself, because the extreme diversity of uses would overwhelm any systematic attempt to impose security standards and ensure that they are respected.  Attempting to do so would probably create some very frustrated users, and yet would at best raise the bar slightly for one facet of a complex, multi-faceted problem.

More interesting, I think, is to think about diversity as a constructive tool for protecting large, complex systems.  Fred Schneider and I once coauthored an editorial on this topic, and a short paper, and I continue to believe that this was one of our better efforts.  Synthetic diversity, in particular, is a remarkable tool for combatting the wily attacker, who often has a surprisingly limited arsenal of off the shelf exploits and might be blocked by even small "surprises".

The basic idea is simple: just as a compiler takes source code and then can create a variety of executables (depending on the target instruction set, operating system, optimization level selected, etc), we can also "compile" programs to expose various forms of artificial diversity.  We can force the memory layout to look somewhat random (simply by adding random padding to objects allocated dynamically).  We can vary the stack layout and the order in which inputs are delivered if a program receives data from multiple sources.  We can potentially compile one program several ways, and pick the version that will be running today at random within the resulting set.  We can randomize the numbering used for operating systems calls.

Such steps diversity and obfuscate the attackable surface.  The attacker who was using an exploit that overruns an input buffer in a way that triggers a system call to download and then execute a bot will now run into an unpredictable memory location, a buffer that might not overflow, and in any case the system call to read from the network socket might have been remapped to some normally-unassigned code.  These are steps that really can pay off.

VPN security offers opportunities too.  Obviously, surfing the web requires that there be no barriers.  But for an application like a medical system that manages patient records or interacts with a lab, there is no real reason to also be able to surface the web, and there is no reason that random web programs should be able to communicate to that application, either.  VPNs can offer this sort of protection, and if we could deploy them more easily, they could wrap sensitive code in layers of armor.

So count me on the side of those who believe that Internet security can and should be a priority, particularly in critical infrastructure and Internet of Things scenarios.   It really is about time to make Internet authentication a readily available and widely standard function, and to deploy VPN technologies much more actively.  But doing so won't solve the whole problem.  We should also make more of an effort to synthetically diversify the systems we deploy widely, so that the attacker will encounter a bewildering variety of "versions" of any particular software.   

If you can't build a single impregnable castle, the next best thing is to populate a huge city with every imaginable variation on the castle theme.  Put police checkpoints on the roads leading to the city water pumping and power generating systems.  Post signs in a few languages, including some you made up just for this purpose.   Good luck to that hacker: he might break into one system, but doing so won't get him very far before we root him out...

Sunday, 11 June 2017

Moving AI to the edge

In today's data centers, a dominant paradigm shapes machine learning and AI systems: massive amounts of data are collected, cleaned, and stored into some form of database or collection of files.  Then machine learning tools are used to train a model on the observed data, and finally the resulting model is used for a while in the edge application.  All of this takes time, so the edge applications operate using stale data, at least to some degree.

In online AI/Ml systems, like smart highways controlling smart cars, smart homes, or the smart power grid, a pervasive need for instant responsiveness based on the most current data is a distinguishing characteristic: today's standard cloud systems can definitely react to new events extremely rapidly (100ms or less is the usual goal), but because the edge runs on cached data -- in this situation, cached models -- and these platforms can't update their models promptly, they will continue to use a stale model long after something fundamental has changed, invalidating it. 

So why would we care about stale models?  The term model, as used by the ML community, refers to any concise representation of knowledge.  For example, on a highway, knowledge of a particular truck's behavior (its apparent route, history of speeds and lane changes, perhaps any observations of risks such as a tire that could be shredding, or a piece of cargo that might not be properly tied down) are all part of the model.  In a continuous learning setting, the model shapes behavior for all the smart cars in the vicinity.  In a smart power grid, the model is our estimate of the state of the grid; if the grid starts to show an oscillatory imbalance or signs of a shortage of power, the model can evolve in milliseconds, and the grid control algorithms need to adjust accordingly.  Yesterday's model, or even the one from ten seconds ago, might not be acceptable.

What will it take to move AI to the edge?
  • The AI/ML community will need to figure out what aspects of their problems really need to run on the edge, and are incorrectly situated on the standard backend today.  Coding at the edge won't be the same as coding for the backend, so work will be required.  This suggests that many applications will need to be split into online and offline aspects, with the online parts kept as slim as possible.  The offline components will be easier to implement because they run in a more standard way.
  • We need programming tools and platforms for edge-oriented AI/ML tools.  I'm betting that Derecho can be the basis of such a solution, but I also think it will take a while to reach that point.  We'll need to understand what edge-hosted AI/ML tools will actually look like: how will they represent, store, and access machine-learned models?  How big are these models, and what data rates arise?  Where are the low-latency paths, and how low does  latency need to be?
  • We may need to integrate hardware accelerators into the infrastructure: if a system does vision, it probably wants to use GPU accelerators for image segmentation, tagging and for operations such as rotation, alignment, debluring, 3-D scene reconstruction, etc.  FPGA components offer a second rapid model, more focused on searching HTML text or other "byte stream" objects.  FPGA accelerators are also useful for doing quick cryptographic operations.  There could easily be other kinds of ASICs too: DFFT units, quantum thinkers, you name it. 
  • All of these pieces need to synchronize properly and be easy to program in a "correct" way...
I'm fascinated by this whole area.  If it interests you too,  send me a note:  perhaps we can find ways to team up!

Monday, 29 May 2017

Byzantine clients

Although I personally don't work on Byzantine block chains, we happen to be in the midst of a frenzy of research and hiring and startups centered on this model.  As you probably know, a block chain is just a sequence (totally ordered) of records shared among some (perhaps very large) population of participating institutions.  There is a rule for adding the next block to the end of the chain, and the chain itself is fully replicated.  Cryptographic protection is used to avoid risk of a record being corrupted or modified after it is first recorded.  A Byzantine block chain uses Byzantine agreement to put the blocks into order, and to force the servers to vote on the contents of every block.  This is believed to yield ultra robust services, and in this particular use-case, ultra robust block chains.

Financial institutions are very excited about this model, and many have already started to use a digital contracts language called hyper-ledger that seems to express a wide range of forms of contracts (including ones that might have elements that can only be filled in later, or that reference information defined in prior blocks), and one can definitely use block chains to support forms of cyber currency, like BitCoin (but there are many others by now).  In fact block chains can even represent actual transfers of money: this is like what the SWIFT international banking standard does, using text messages and paper ledgers to represent the transaction chain.

Modern block chain protocols that employ a Byzantine model work like this: we have a set of N participants, and within that set, we assume that at most T might be arbitrarily faulty: they could collude, might violate the protocol, certainly could lie or cheat in other ways.  Their nefarious objective might be to bring the system to a halt, to confuse non-faulty participants about the contents of the blockchain or the order in which the blocks were appended, to roll-back a block (as a way to undo a transaction, which might let them double-spend a digital coin or renege on a contract obligation), etc.  But T is always assumed to be less than N/3.  

Of course, this leads to a rich class of protocols, very well-studied, although (ironically), more often in a fully synchronous network than in an asynchronous one.  But the so-called PRACTI protocols created by Miguel Castro with Barbara Liskov work in real-world networks, and most block chain systems use some variation of them.  They basically build on consensus (the same problem solved by Lamport's Paxos protocol).  The non-compromised participants simply outvote any Byzantine ones.

What strikes me as being of interest here is that most studies have shown that in the real world, Byzantine faults almost never arise!   And when they do, they almost never involve faulty servers. Conversely when attackers do gain control, they usually compromise all the instances of any given thing that they were able to attack successfully.  So T=0, or T=N.

There have been a great many practical studies on this question.  I would trace this back to Jim Gray, who once wrote a lovely paper on "Why Do Computers Stop and What Can Be Done About It?".  Jim was working at Tandem Computers at that time, on systems designed to tolerate hardware and software problems.  Yet they crashed even so.  His approach was to collect a lot of data and then sift through it.

Basically, he found that human errors, software bugs and design errors were a much bigger problem then hardware failures.  Jim never saw any signs of Byzantine faults (well, any fault is Byzantine.  But I mean malicious behaviors, crafted to compromise a system).

More recent studies confirm this.  At Yahoo, for example,  Ben Reed examined data from a great many Zookeeper failures, and reported his findings ina WIPS talk at SOSP in 2007.  None were Byzantine (in the malicious sense).

At QNX Chris Hobbs has customers who worry that very small chips might experience higher rates of oddities that would best be modeled as Byzantine.  To find out, he irradiated some chips in a nuclear reactor, and looked at the resulting crashes to see what all the bit flips did to the code running on them (and I know of NASA studies of this kind, too).  In  fact, things do fail.  But mostly, by crashing.  The main issue turns out to be undetected data corruption, because  most of the surface area on the chips in our computers is used for SRAM, caching, and DRAM  storage.  Data protected by checksums and the like remains safe, but these other forms are somewhat prone to undetected bit flips.  But most data isn't in active use, in any case: most computer memory just has random stuff in it, or zeros, and the longest lived active form of memory is to hold the code of the OS and the application programs.  Bit-flips will eventually corrupt instructions that do matter, but corrupted instructions mostly trigger faults.  So, it turns out that the primary effect of radiation is to raise the rate of sudden crashes.

Back when Jim did his first study, Bruce Nelson built on it by suggesting that the software bugs he was seeing fall into two cases: Bohrbugs (deterministic and easily reproduced, like Bohr's model of the atom: an easy target to fix) and Heisenbugs (wherever you look, the bug skitters off to somewhere else).  Bruce showed that in a given version of a program, Bohrbugs are quickly eliminated, but patches and upgrades often introduce new ones, creating an endless cycle.  Meanwhile, the long-lived bugs fell into the Heisenbug category, often originating from data structure damage "early" in a run that didn't cause a crash until much later, or from concurrency issues sensitive to thread schedule ordering.  I guess that the QNX study just adds new elements to that stubborn class of Heisenbugs.

So, we don't have Byzantine faults in servers, but we do have a definite issue with crashes caused by bugs, concurrency mistakes, and environmental conditions that trigger hardware malfunctions. There isn't anything wrong with using Byzantine agreement in the server set, if you like.   But it probably won't actually make the service more robust or more secure.

Bugs can be fought in many ways.  My own preferred approach is simpler than running Byzantine agreement.  With Derecho or other similar systems, you just run N state machine replicas doing whatever the original program happened to do, but start them at different times and use state transfer to initialize them from the running system (the basis of virtual synchrony with dynamic group membership).

Could a poison pill kill them all at once?  Theoretically, of course (this is also a risk for a Byzantine version).  In practice, no.  By now we have thirty years of experience showing that in process group systems, replicas won't exhibit simultaneous crashes, leaving ample time to restart any that do crash, which will manifest as occasional one-time events.

Nancy Leveson invented a methodology called N-Version programming; her hypothesis was that if most failures were simply due to software bugs, then that by creating multiple versions of the most important services, you could mask the bugs because coding errors would probably not be shared among the set.  This works, too, although she was surprised at how often all N versions were buggy in the same way.  Apparently, coders often make the same mistakes, especially when they are given specifications that are confusing in some specific respect.

Fred Schneider and others looked at synthetic ways of "diversifying" programs, so that a single piece of code could be used to create the versions: this method automatically generates a bunch of different versions from one source file, with the same input/output behavior.  You get N versions too, often find that concurrency problems won't manifest in the identical way across the replicas, and they also are less prone to security compromises!

Dawson Engler pioneered automated tools for finding bugs and even for inferring specifications.  His debugging tools are amazing, and with them, bug-free code is an actual possibility.

Servers just shouldn't fail in weird, arbitrary ways anymore.  There is lots of stuff to worry about, but Byzantine compromise of a set of servers shouldn't be high on the radar.

Moreover, with strong cloud firewalls, Intel SGX, and layers and layers of monitoring, I would bet that the percentage of compromises that manage to take control of between and N/2-1 replicas (but not more), or. even of attacks that look Byzantine is even lower.

But what we do see  (including Reed's 2007 study), are correct services that come under attack from some form of malicious client.  For an attacker, it is far easier to hijack a client application than to penetrate a server, so clients that try to use their legitimate  connectivity to the server for evil ends are a genuine threat.  In fact with open source, many attackers just  get the client source code, hack it, then run the compromised code.  Clients that simply send bad data without meaning to are a problem too.

The client is usually the evil-doer.  Yet the Byzantine model blames the server, and ignore the clients.

It seems to me that much more could be done to characterize the class of client attacks that a robust service could potentially repel.  Options include monitoring client behavior and blacklisting any clients that are clearly malfunctioning, capturing data redundantly and somehow comparing values, so that a deliberately incorrect input can be flagged as suspicious or even suppressed entirely (obviously, this can only be done rarely, and for extreme situations), or filtering client input to protect against really odd data.  Gun Sirer and Fred Schneider once showed that a client could even include a cryptographic  proof that input strings really came from what the client typed, without tampering.

Manuel Costa and Miguel Castro came up with a great system they called Vigilante, a few years ago.  If a server was compromised by bad client input, it detected the attack,  isolated the cause and spread the word instantly.  Other servers could then adjust their firewalls to protect themselves, dynamically.  This is the sort of thing we need.

So here's the real puzzle.  If your bank plans to move my accounts to a block chain, I'm ok with that, but don't assume that BFT on the block chain secures the solutions.  You need to also come up without a way to protect the server against buggy clients, compromised clients, and clients that just upload bad data.  The "bad dudes" are out there, not inside the data center.  Develop a plan to to keep them there!

Friday, 26 May 2017

More thoughts on the learnability of implicit protocols

While driving from Sebastopol (in Sonoma valley, north of San Francisco) to Silicon Valley (which is south of San Francisco), one has a lot of time to think about how human-operated cars coordinate and implicitly communicate.  So I thought I might follow up on the posting I did a few weeks ago concerning the learnability of implicit protocols and see if I can't make the question a bit more concrete.

The best way to simplify a problem is to think of a base case, so I started by asking myself about the simplest model I could: vehicles modeled as points and a highway modeled as a straight line.

Interactions between vehicles would basically define a graph: any given car gets a chance to interact with the car in front of it, and the one behind it, and the protocol (whatever it might be) is like a state machine on these graphs.

So what behaviors would constitute a protocol in such a situation?  Well, we presumably would have car A in state Sa, and car B in state Sb, and then some "communication" takes place.  Not too many options: they either draw closer to one-another at some speed, move apart from one-another at some speed (I mean speed of closing or speed of separating, but probably the joint speed and direction of movement needs to be specified too), or perhaps one or the other brakes sharply or accelerates sharply.  So here we have a "vocabulary" for the cars, with which they communicate.

A protocol, then, would be a rule by which A reacts to inputs of this form from B, and vice versa, since any action by B on A is also perceived by B itself (that is, even when B probes A, B itself may be forced to take some reactive action too: the human version is that you are trying to shift from some horrendously bad radio station to a different, slightly less horrible, alternative when you look up and notice that you've gotten much closer to the car in front you than you expected.  So in this case it was your own darn fault, but just the same, you react!)

So to learn a protocol, we might imagine a series of "tests" by which vehicle B, approaching A from behind, experiments to investigate A's reaction to various inputs, recording these into a model that over time would converge towards B's approximation of A's algorithm.  A similarly is learning B's behavior (and A might elicit behaviors from B.  For example, sometimes a car comes up on my tail very close, and especially if we are already basically moving at the fastest possible speed, it annoys me enough that I gradually slow down, rather than speeding up, as he or she probably intended for me to do -- there is an optimal distance at which to follow me, if you are trying to do so.  Similarly for most drivers.  Drive 3x further back and I might ignore you completely.  Drive at half that distance and my behavior departs from the norm: I slow down, which is suboptimal for both of us.  My way of punishing your pushy behavior!).

If you could figure out how two cars can learn a protocol, then you could ask if this generalizes to situations that form trains of cars.  So here, B catches up with a train: not just A, but perhaps X-Y-Z-A, with A in the rear and some form of agreed-upon protocol in use by X,Y,Z and A.  B gets to join the train if it can learn the protocol and participate properly.  Failing to do so leaves B abandoned, or results in an accident, etc.

Optimal behavior, of course, maximizes speed and minimizes the risk of accidents.  I'm starting to see how one could add a kind of utility function here: B wants to learn A's protocol quickly and "efficiently", with as few tests as possible.  Then B wants to use the knowledge gained to achieve this highest possible speed "jointly" with A.  That defines a paired protocol, and the generalization becomes a train protocol.

Another observation: Suppose that B has a small set of models in a kind of grab-bag of models that cars commonly use.  Now B's job of learning the model is potentially easier: if A is using one of the standard models, it may only take a very small number of tests to figure that out.  Moreover, having a good guess might make some models "learnable" that would actually not be learnable with a small number of tests otherwise.  This would fit well with the experience I've had in New Jersey, Belgium, Tel Aviv and California: each country has its own norms for how closely cars tend to space themselves on high speed throughways and in other situations, so while you can feel disoriented at first, in fact this only adds up to four models.  If someone handed you those four models, I bet you could figure out which one applies with very few experiments.

Unless, of course, you accidentally end up behind a New Jersey driver who is learning the rules of the road in Tel Aviv for the first time, of course.  I think I've encountered a few such cases too.

So that would be a further interesting case to consider: roads with mixtures of models: subsets of cars, with each subset using a distinct model.  In fact my story of New York City cab drivers, from last time, would fall into that category.  The cabs figure out how to form school-of-fish packs that flow around obstacles and other cars, no matter what those other cars might be doing.  Meanwhile, there could easily be other protocols: New York public transit buses, for example, have their own rules (they basically are bigger than you, and can do whatever they like, and that is precisely how they behave).

Conversely, the New Jersey driver also illustrates a complication: the implicit protocols of interest to me are cooperative behaviors that maximize utility for those participating in the protocol, and yield higher utility than could be gained from a selfish behavior.  But any vehicle also will have an autonomous set of behaviors: a protocol too, but one that it engages in purely for its own mysterious goals.  And these selfish behaviors might not be at all optimal: I recently watched a teenage driver in the car behind me putting on makeup while talking on a cell phone and apparently texting as well.  So this particular driver wasn't maximizing any normal concept of utility!  And all of us have experienced drivers in the grips of road-rage, driving hyper aggressively, or teenage guys hot-dogging on motorcycles or weaving through traffic in their cars.  Their sense of utility is clearly very warped towards speed and has a very diminished negative utility for things like accidents, traffic violations and tickets, or sudden death.  So even as we learn to cooperate with vehicles prepared to cooperate, and need to figure out the protocols they are using, the environment is subject to heavy noise from these kinds of loners, inattentive drivers, reckless drivers, and the list goes on.

One thing that strikes me is that the cab driver protocol in New York, once you know it, is a good example of a very easily discovered policy.  Despite a huge level of non-compliant vehicles, anyone who knows the cab protocol and drives using it will quickly fall into synchrony with a pack of others doing so.  So the level of positive feedback must be massive, if we had the proper metric, relative to the level of noise. 

Interestingly, there is clearly a pack-like behavior even for a line of cars driving on a single lane.  This would be a very good case to start by understanding.

Of course the real-world-school-of-fish behaviors involve (1) shifting lanes, which is the easy case and (2) ignoring lane boundaries, which is a harder case.  Given that even the cars on a string case taxes my imaginative capabilities, I think I'll leave the full 2-dimensional generalization to the reader!