Home » 2018
Yearly Archives: 2018
MICA: A Holistic Approach to Fast In-Memory Key-Value Storage
Hyeontaek Lim, Dongsu Han, David G. Andersen, and Michael Kaminsky in Proceedings of the 11th USENIX Symposium on Networked Systems Design and Implementation (NSDI ’14). April 2-4, 2014, Seattle, WA.
It’s time to turn attention to key value stores. There are quite a few papers on this subject. Seemingly simple, they turn out to be an important storage mechanism. I had a delightful discussion about them over the weekend with Keith Bostic, who knows a fair bit about key-value stores (being the author of one of the most widely used KV stores.) Of course, I have discussed some KVS already: MassTree, NVMCached, and SILT.
I’m going to continue looking at key-value stores (KVS) by starting with MICA. It is neither the first, nor the latest, work in this area, but it seems like a good place to start really describing key value stores in this space. I have quite a few to cover so let’s get started.
MICA is a scalable in-memory key-value store that handles 65.6 to 76.9 million key-value operations per second using a single general-purpose multi-core system. MICA is over 4–13.5x faster than current state-of-the-art systems, while providing consistently high throughput over a variety of mixed read and write workloads.
It is an interesting system in that the authors decided to look at making their KVS tunable – it can serve as a cache (which means items can be thrown out by the KVS as space becomes scarce) or it can serve as storage (which means items cannot be removed from the KVS by the system itself, though it permits the applications using it to do so.)
To achieve this, they propose a series of data structures tuned to the task: circular logs, lossy concurrent hash indexes, and bulk chaining. Their goal is to provide high performance through low contention. They utilize dpdk for fast network processing. They split their data structures so they independently handle caching and storage separately. I might argue that this makes it two systems fused together, but let’s give them the benefit of the doubt at this point by calling it “generality”.
One of the concerns they try to address is the hot key issue. In many KVS systems they get terrific performance on synthetic workloads because they use a uniform distribution. In the real world, particularly for web objects, the workloads tend to be mostly cold with a small subset of the keys that are hot. Further, these keys change over time.
They do restrict their domain somewhat to make the problem more tractable. For example, they do not handle large items – everything must fit within a single network packet. For systems that require larger items, they suggest using a separate memory allocator. Their argument is that the extra indirection cost is a marginal increase in total latency. They are optimizing for the common case. Another aspect they exclude is durability. That makes sense – they are only implementing an in-memory KVS, so durability is not a logical part of their work. This also plays well in terms of performance, since persistence makes things much more complicated.
So what do they want to achieve with MICA?
- High single-node throughput
- Low end-to-end latecy
- Consistent performance across workloads
- Support small (but variable length) key-value objects
- Support basic KVS operations: get, put, delete
- Work efficiently on commodity hardware
They propose achieving this using a combination of design choices:
- Parallel data access – in fact, they mean read access, as they do not parallelize these, as the cost is too high.
- Optimized network overhead – as noted previously, they use dpdk to achieve this
- Efficient KV structures – this includes dual memory allocators (one optimized for caching the other for store) and efficient index implementations.
There are some interesting characteristics to this system, including the fact that it employs dangling pointer handling. Rather than remove pointers from the index, they allow the pointers to become invalid and instead trap that condition. The paper describes how they do that in more detail.
They also use an interesting “prefetcher” to encourage efficient memory loading. This is done by receiving requests in bursts and scheduling those requests to be processed “soon” to this prefetch step.
Their evaluation of the system tries to focus on a range of workloads, varying key and value size as well as the range of operations being performed, with a read intensive workload being 95% reads and 5% writes. A write intensive workload is split evenly.
They also look at how the key space is handled. In the “Exclusive Read, Exclusive Write” paradigm (EREW) each core of the system is responsible for handling both reads and write for a given key (or range of keys more likely). In the “Concurrent Read, Exclusive Write” paradigm (CREW), any core may satisfy a read for the key, but only one core may satisfy a write operation for the core. This is a hat tip towards the cost of “bouncing” control of the cache lines between cores and we will see this approach used by other systems as well. One interesting question that this reminds me of each time I read about this model is what impact, if any, the Level 3 cache might have on this. Level 1 and 2 caches are traditionally per core but Level 3 is typically per socket. Thus, things stored in L3 cache are substantially faster to access than RAM. Something to consider, perhaps.
The paper has an extensive evaluation section and, of course, they demonstrate how their solution is substantially faster. One interesting aspect to their evaluation is that they claim that each component of their design decision is important in enabling them to achieve their target performance. To do this, they evaluate each individual aspect of their design decision. For example for their parallel data model, they evaluate the variations on their choice and demonstrate that CREW is only faster with read skewed workloads. They even go so far as to model Concurrent Read Concurrent Write (CRCW) workloads and point out that the cache overhead blunts the benefits of the concurrency.
To make this point on their choice of network implementations, they take MassTree and convert it to use dpdk, which yields a substantial performance improvement. While they argue that their key indexes also contribute, they do not have a distinct break-out of that implementation and fall back to pointing out how they work better, even in CREW than MassTree.
There are certainly grounds to criticize MICA. For example, they do choose workloads that fit well with their available resources. I suspect that once they become resource constrained (e.g., cache) that the performance will drop substantially. However, it provides an interesting contribution in the space and does point out how picking your data structures carefully can make a tremendous difference in the overall performance of the system.
Message Passing or Shared Memory: Evaluating the Delegation Abstraction for Multicores
Irina Calciu, Dave Dice, Tim Harris, Maurice Herlihy, Alex Kogan, Virendra Marathe, and Mark Moir, in International Conference on Principles of Distributed Systems, pp. 83-97, 2013, Springer.
The issue of isolation versus sharing is one that permeates systems design for the past 50 plus years. Isolation makes things clean and easy to reason about, but suffers from some disadvantages, such as increased resource utilization and increased latency costs for communications. Shared memory is actually much more difficult to program due to the need to coordinate activity to a shared resource (the memory) potentially across processors. In this paper the authors explore this space. Their rationale is simple: shared memory with lots of processors is
Even for small multi-core systems, it has become harder and harder to support a simple shared memory abstraction: processors access some memory regions more quickly than others, a phenomenon called non-uniform memory access (NUMA). These trends have prompted researchers to investigate alternative programming abstractions based on message passing rather than cache-coherent shared memory. To advance a pragmatic understanding of these models’ strengths and weaknesses, we have explored a range of different message passing and shared memory designs, for a variety of concurrent data structures, running on different multicore architectures. Our goal was to evaluate which combinations perform best, and where simple software or hardware optimizations might have the most impact. We observe that different approaches per-form best in different circumstances, and that the communication over-head of message passing can often outweigh its benefits. Nonetheless, we discuss ways in which this balance may shift in the future. Overall, we conclude that, by emphasizing high-level shared data abstractions, software should be designed to be largely independent of the choice of low-level communication mechanism.
Non-uniform memory access (NUMA) was originally constructed as a scaling mechanism for “large scale systems”. Today, it is generally present on any multi-processor system. What this means is that memory is divided between the processors. Some of the memory is bound to the processor and then the rest of the memory is bound to other processors. Thus, memory is either local or remote, with a corresponding cost differential to access it.
Each processor usually has multiple cores, with each of those cores making up one element of the node. Some computers may have multiple sockets within a single NUMA node. No doubt there are other models for managing physical memory beyond is basic model, but the idea is the same.
The authors are thus observing that this configuration has become so common that it is now commonly observed on real world systems. This in turn means that programmers must deal with this. One way to achieve this is to change the tools so they hide this complexity from the typical programmer; for some this works fine. For operating systems and high performance systems, such as key-value stores (KVS), it is important to understand these issues in order to optimize their performance.
So these authors look at performance across a “range of message passing and shared-memory designs”. This sounds like a laudable goal to me.
They set out to define things. Delegation is where access to a data structure is only performed by a specific set of threads. If another thread wishes to make a modification to the data structure, it must request that one of the controlling threads do so by sending a request. A controlling thread for the data structure validated the operation, performs it, and then returns the results to the original requester. Simple. In addition, it can be used to simplify the locking model – after all, if a single thread controls the data structure then there cannot be any contention for the data structure and thus there is no locking required. In addition, if that controlling thread is running on one core and the requester thread is running on a different processor, you can end up with better resource utilization, notably the processor caches.
But the downside to delegation is that you introduce a message passing delay. If the data structure is hot – because it has a high volume of usage – then the delegation mechanism can become a bottleneck, involving queuing delays.
The authors then turn their attention to message passing overhead. They evaluate different queuing mechanisms, though their message format remains fairly uniform. Their queues are the multiple producer, single consumer (MPSCChannel), a per-NUMA node message queue (InletQueue), and a direct access without using compare-and-swap queue (DNCInletQueue). They carefully describe how each is implemented. Then they turn to shared memory, where the issue is synchronizing access to the shared memory. They use several different locking mechanisms: a spin lock, an MCS lock (“ticket lock”), and two variations of a NUMA optimized MCS lock, one that attempts some level of fairness and the other that does not.
Then given these simple mechanisms they evaluate against a concurrent hash map and a concurrent linked list. They describe their memory allocation strategy (since they are storing key/value data). The use a two key workloads (“small” and “large”) and then use three operations mixes: a get-only mix, an insert/delete only mix, and a 50/50 mix of read and write operations.
Then they describe their results: small hash maps just don’t benefit from more threads. Delegation is slower than shared memory. Simple MCS locks do best under minimal contention. Interestingly, unfair NUMA aware MCS locks perform best under high contention of the four lock types. Linked lists is where delegation wins. The authors argue this is due to better cache locality. They point out considerable variance in the ticket locks depending upon the workload.
The also evaluated a mixed workload – trying to see the “hot spot” problem in their evaluation. Once again their unfair NUMA aware ticket lock has the best performance, but the authors point out that it introduces substantial delay for some threads. If you are concerned about bounded latencies, this is not the lock to use.
They had an interesting observation about the impact of hyper-threading on performance (“sibling rivalry”) and end up ensuring that they do not compete for resources on the same core between clients and servers for the delegation case.
They point out that reducing contention for locking is important because it helps minimize the impact on the overall memory bus; thus minimizing cache contention is helpful to the overall system performance, not just the specific applications in question.
Their conclusion? “[D]elegation can sometimes outperform direct shared-memory approaches”.
We’re going to see some strong arguments for this position in later papers.
NVMCached: An NVM-based Key-Value Cache
Xingbo Wu, Fan Ni, Li Zhang, Yandong Wang, Yufei Ren, Michel Hack, Zili Shao, and Song Jiang, in Proceedings of the 7th ACM SIGOPS Asia-Pacific Workshop on Systems, 2016.
This is a short (6 page) paper that discusses the challenges of taking a memory based key-value cache (e.g., memcached) and converting it to use a non-volatile memory. Because the authors do not have NVM available, they focus on simulating it using DRAM. They simulate behaviors by using appropriate cache flushes and fence. In fact, they use CLFLUSH and MFENCE, which may be a bit more than is required.
Here is an interesting thing for people familiar with Intel processor architecture for more than about the past 15 years: the CPUs no longer provide strong ordering guarantees. If you’ve learned or refreshed your knowledge about the memory consistency models more recently, then you are likely already familiar with this. One of the interesting issues with NVM is that we need to think about memory consistency in a way that’s never been required previously. The authors of this paper do this.
They note that flushes are expensive; they argue for using “consistency-friendly data structures” for minimizing the number of flushes required, which certainly makes sense, particularly given that they are expensive. Some operations (like anything with the LOCK prefix) are performed atomically as well and are conceptually similar; both tie into the processor’s cache architecture (for those that have been around a while, the cache line semantics for LOCK were introduced around 1995.)
One of the important observations in this paper is that the reason to use NVM is that it allows fast restart after a system failure, such as a crash or power event. For memcached, there are users that are asking for persistence. Even if the service is only a cache, it turns out that offloading to the cache can be a powerful mechanism for improving overall throughput, but the larger the cache becomes, the higher the cost when the cache gets tossed after a system crash.
However, this also simplifies the problem at hand: discarding data in this type of system is of minimal concern. They utilize checksums to detect damaged data structures, which alleviates their need to perform ordered write operations; they reorganize their data structures to permit changes with only infrequent fencing. They point out the problem of zombie data (data that is deleted, then the system crashes, so the data returns after the system reboots). To prevent this situation they choose to split their data structures.
The checksums prove to be much less expensive than flush; the authors claim up to 20x faster (my own reading indicates that using the optimized CRC32 built into modern Intel CPUs the cost is just over 1 clock cycle per byte checksummed, so a 64 byte cache line would have a cost of approximately 70 clock cycles; on a 2.0GHz processor, that would be roughly 17.5 nanoseconds, which certainly is somewhere between 5 and 20 times faster than writing back a cache line.
The authors also touch on the challenges of memory allocation. Memory fragmentation is something we often fix in dynamic memory by rebooting the machine. However, now we have the storage level problem with fragmentation; fixing this involves write amplification. They organize their allocation strategy by observing there is considerable locality across the KV store. They combine a log structured allocation scheme with a “zone based” allocation model and attempt to combine cleaning, eviction, and updates to minimize the impact on performance. They also use DRAM to track access information as part of their LRU implementation. This creates a burden on crash recovery, but benefits from significant improved runtime performance.
They also implement a DRAM write combining cache; the idea is that when an item is detected to be “hot” the item is cached in DRAM. Those hot items will thus be discarded if the system crashes.
Their evaluation is against a modified version of memcached. They utilize four Facebook traces; interestingly they do not use one (VAR) because “it is write-dominant with only a few distinct keys touched.” I am not certain that I see why this makes it unsuitable for use, though I suspect they do not obtain favorable behavior in the face of it.
They evaluate the performance of native memcached versus their NVM-enhanced version with a series of micro-benchmarks after a restart event. They show that in several cases it takes more than 1 billion requests before the performance of the DRAM memcached to converge with the NVM enhanced version. When they use the “real world” Facebook memached traces, they observe that three of the benchmarks show comparable performance. Two of them, SYS and VAR, show substantially better performance. Both benchmarks are write intensive. I admit, this surprised me, since I am not certain why writing to (simulated) NVM should be substantially after than writing to DRAM.
Overall, it is interesting work and touches on important challenges for working with NVM as well as improving system performance.
The Future of Synchronization on Multicores: The Multicore Transformation
Maurice Herlihy in Ubiquity, September 2014.
I’m going to round out the week with a much lighter read. Despite this, it has some useful observations that underlie some of the other papers that I’ve been discussing.
The editor’s introduction to this piece really does a good job of summing up the problem:
Synchronization bugs such as data races and deadlocks make every programmer cringe — traditional locks only provide a partial solution, while high-contention locks can easily degrade performance. Maurice Herlihy proposes replacing locks with transactions. He discusses adapting the well-established concept of data base transactions to multicore systems and shared main memory.
The author points out: “Coarse-grained locks … generally do not scale: Threads block one another even when they do not really interfere, and the lock itself becomes a source of contention.” I have personally experienced this and moved on to the next solution, which has its own separate problems: “Fine-grained locks can mitigate these scalability problems, but they are difficult to use effectively and correctly.”
When I have taught about locking in the past, I’ve often approached it from the debugging perspective: fine-grained locks create deadlocks, which can be almost impossible to debug without instrumentation. In operating systems, we prevent deadlocks by defining a lock hierarchy. The order in which locks can be acquired forms a graph. To prevent deadlocks, we require that the graph be acyclic. That sounds simple and for simple code bases, it is. However, in the real world where we introduce such fine-grained locks, the code base is seldom simple and we end up finding complex situations, such as re-entrant behavior, where the cycles appear. Cycles can be introduced because we have multiple discreet components, each doing something logical, that creates a lock cycle unwittingly.
The author also points out another problem with locks that is important in real systems: “Locks inhibit concurrency because they must be used conservatively: a thread must acquire a lock whenever there is a possibility of synchronization conflict, even if such conflict is actually rare.” A common maxim in systems programming is to optimize the common case. Locks do the opposite: they burden the common case with logic that is normally not useful.
The author also points out that our lock mechanisms do not compose well: when we need to construct consistent higher level logic from lower level locked primitives, we have no simple way to interlock them unless they expose their own locking state. I have built such systems and the complexity of verifying state after you acquire each lock and unwinding when the state has changed is challenging to explain and conceptualize.
This is so complicated that in many cases concurrency is handled within the tools themselves in order to insulate the programmer from that complexity. It may be done by isolating the data structures – single threaded data structures don’t need locks – so you can use isolation and message passing. It can be done in a transactional manner, in which the locking details are handled by the tools and lock issues cause the transaction to roll back (abort), leaving the application programmer to restart again (or the tools to attempt to handle it gracefully).
One such way to achieve this is to implement transactional memory: a series of operations that are performed sequentially and once the operation is done, the outcome is determined: the transaction either becomes visible (it is committed) or it fails (it is aborted) and no changes are made. General transaction systems can be quite complicated: this is a common database approach.
How do we make transactions simple enough to be useful in multicore shared memory environments?
- Keep them small: they don’t change much state
- Keep them brief: they either commit or abort quickly
- Keep them ephemeral: they don’t involve disk I/O, they aren’t related to persistence they are related to consistency.
One benefit of transactions is they are composable: they can be nested. Transactions can avoid issues around priority inversion, convying, and deadlocks. The author points to other evidence that says they’re easier for programmers and yield better code.
Transactions aren’t new. We’ve been using them for decades. When we use them at disk I/O speeds, we find the overhead is acceptable. When we use them at memory speeds we find the overhead of transactions is too high to make them practical to do in software. This gave birth to the idea of hardware transactions. Hardware transactions can be used in databases (see Exploiting Hardware Transactional Memory in Main-Memory Databases) quite effectively. They don’t suffer from the high overhead of software transactions. The author points out a limitation here: “Hardware transactions, while efficient, are typically limited by the size and associativity of the last-level cache”. When a cache line cannot remain in the CPU, the transaction is aborted. Software must then handle the abort: “For these reasons, programs that use hardware transactions typically require a software backup.” As we saw in previous work (again Exploiting Hardware Transactional Memory in Main-Memory Databases) just retrying the operation once or twice often resolve the fault. But sometimes the operation is just not viable on the system at the present time.
The author’s summary of the impact of hardware transactions is interesting:
The author predicts that direct hardware support for transactions will have a pervasive effect across the software stack, affecting how we implement and reason about everything from low-level constructs like mutual exclusion locks, to concurrent data structures such as skip-‐lists or priority queues, to system-‐level constructs such as read-‐copy-‐update (RCU), all the way to run-time support for high-‐level language synchronization mechanisms.
So far, this change has not been pervasive. I have seen signs of it in the operating system, where lock operations now take advantage of lock elision in some circumstances. Systems, and software stacks, do change slowly. Backwards compatibility is a big issue. As we move forward though, we need to keep these new mechanisms in mind as we construct new functionality. Better and faster are the goals.
Exploiting Hardware Transactional Memory in Main-Memory Databases
Viktor Leis, Alfons Kemper, Thomas Neumann in 2014 IEEE 30th International Conference on Data Engineering, pp 580-591.
I have not spent much time discussing transactional memory previously, though I have touched upon it in prior work. By the time this paper was presented, transactional memory had been fairly well explored from a more theoretical perspective. Intel hardware with transactional memory support was just starting to emerge around the time this paper was released. I would note that Intel had substantial challenges in getting hardware transactional memory (HTM) correct as they released and then pulled support for it in several different CPU releases. Note that HTM was not new, as it had been described in the literature to an extent that earlier papers (e.g., Virtualizing Transactional Memory, which I have decided not to write about further, discusses the limitations of HTM back in 2005).
Logically, it extends the functionality of the processor cache by tracking what is accessed by the processor (and driven by the program code). Cache lines are read from memory, any changed made to the cache line, and then written back to memory. This is in turn all managed by the cache coherency protocol, which provides a variety of levels of coherency.
The idea behind HTM is that sometimes you want to change more than a single element of memory. For example, you might use a mutual exclusion, then add something to a linked list, and increment a counter indicating how many elements are in the linked list before you release the mutual exclusion. Even if there is no contention for the lock, you will pay the lock cost. If the platform requires a fence operation (to ensure memory has been flushed properly) you will also stall while the memory is written back. In a surprising number of cases, you need to do multiple fences to ensure that operations are sequentially consistent (which is a very strong form of consistency).
With HTM you can do this all speculatively: start the transaction, add something to the linked list, increment the counter, then commit the transaction. Once this has been followed with an appropriate fence, the change is visible to all other CPUs in the system. The goal then is to avoid doing any memory operations unless absolutely necessary.
The authors point out that the fastest option is partitioning (ignoring hot spots). They graphically demonstrate this in Figure 1 (from the paper). HTM has some overhead, but it tracks with partitioning fairly linearly. This difference is the overhead of HTM.
They compare this to serial execution, which just means performing them one at a time. The traditional mechanism for doing this kind of paralleism is the two phase commit protocol. That’s the lock/work/unlock paradigm.
If we only considered this diagram, we’d stick with strong partitioning – and we’re going to see this observation reflected again in future work. Of course the reason we don’t do this is because it turns out that the database (and it shows up in file systems as well) is not being uniformly accessed. Instead, we have hot spots. This was a particular concern in the MassTree paper, where they supported novel data structures to spread the load around in a rather interesting fashion. There’s quite a bit of discussion about this problem in the current paper – “[A] good partitioning scheme is often hard to find, in particular when workloads may shift over time.” Thus, their observation is: “we have to deal with this problem”.
So, how can HTM be exploited to provide robust scalability without partitioning. The authors do a good job of explaining how HTM works on Intel platforms. Figure 4 (from the paper) shows a fairly standard description of how this is done on the Intel platform: it has a bus snooping cache, an on-chip memory management unit (MMU), a shared Level 3 cache, and per core Level 1 and Level 2 caches (in case you are interested, the two caches do have somewhat different roles and characteristics.) Level 1 cache is the fastest to access, but the most expensive to provide. Level 2 cache is slower than Level 1, but because it is also cheaper we can have more of it on the CPU. Level 3 cache might be present on the CPU, in which case it is shared between all three cores. Note that none of this is required. It just happens to be how CPUs are constructed now.
The benefit of HTM then is that it exploits the cache in an interesting new way. Changes that are made inside a transaction are pinned inside the cache so they are not visible outside the current core. Note, however, that this could mean just the L1 cache. In fact, the functional size permitted is even smaller than that, as shown in Figure 5 (from the paper). Transactions below 8KB have a low probability of aborting (and if it aborts, the operation failed so it must be tried again, either using HTM or the fallback mechanism with software). That probability approaches 100% as the size goes above above 8KB. Interestingly, the primary reason for this is not so much the size of the cache as the associativity of the cache. What that means is the cache uses some bits from the address to figure out where to store data from that particular cache line. The paper points out that 6 bits (7-12) are used for determining the cache location, and each cache location (so each unique value of bits 7 through 12) are has a fixed number of cache lines (e.g., 8 entries in the Haswell chips the authors are evaluating). If we need to use a ninth we evict one of the existing pages in the cache.
Similarly, when the duration of the transaction goes up, the probability of it aborting also rises. This is shown in Figure 6 (from the paper). This is because the chance that various systems events will occur, which cause the transaction to abort. This includes various types of interrupts: hardware and software.
We also note that we should take steps to minimize the amount of sharing of data structures that might be required – the point that not sharing things is more efficient. The authors discuss a variety of approaches to this issue: segmenting data structures, removing unnecessary conflict points (e.g., counters), and appropriate choice of data structures.
Recall the Trie structures from MassTree? These authors offer us Adaptive Radix Trees, which seem to have a similar goal: they are “[A]n efficient ordered indexing structure for main memory databases.” They combine this with a spin lock; the benefit now is that HTM doesn’t require the spin lock normally, so even if some parts of the tree are being read shared, the lock is not being acquired and thus it does not force a transactional abort for other (unrelated) nodes.
They put all of this insight together and that forms the basis for their evaluation. Figure 11 in the paper makes the point that HTM scales much better than traditional locking for small lookups (4 byte keys) with a uniform distribution once there is more than one thread.
Figure 12 (from the paper) evaluates the TPC-C Benchmark against their resulting system to demonstrate that it scales well . Note they stick with four threads, which are all likely on a single physical CPU, so there are no NUMA considerations in this aspect of the evaluation. They address this a bit later in the paper.
Figure 13 (from the paper) compares their performance against a partitioned system. Because they cannot prevent such cross-partition access, they must “live with” the inherent slowdown. One of the amazing benefits of HTM is thus revealed: as more operations cross partition boundaries, HTM continues to provide a constant performance. This seems to be one of the key lessons: no sharing is great, but once you find that you must share, synchronizing optimistically works surprisingly well.
Figure 14 (from the paper) attempts to address my comment earlier abut Figure 12: they really don’t have a multiprocessor system under evaluation. They admit as much in the paper: the hardware just isn’t available to them. They provide their simulation results to defend their contention that this does continue to scale, projecting almost 800,000 transactions per second with 32 cores.
Figure 15 (from the paper) finally demonstrates the reproducibility of HTM abort operations. If an HTM is retried, many will complete with one or two tries. Thus, it seems that even with multiple threads, they tend to converge towards the hardware limitations.
Bottom line: hardware transactional memory can be a key aspect of improving performance in a shared memory systems with classical synchronization.
SILT: A Memory-Efficient, High-Performance Key-Value Store
Hyeontaek Lim, Bin Fan, David G. Andersen, Michael Kaminsky, in Proceedings of the 23rd Symposium on Operating Systems Principles (SOSP ’11), October 23-26, 2011, Cascais, Portugal.
In this paper, we have an interesting hybrid key-value store; part of the store is in DRAM, while the rest of it is stored in flash storage. The authors have focused on optimizing the performance of the store across the specific media chosen, which leads to some interesting choices and observations.
SILT (Small Index Large Table) is a memory-efficient, highperformance
key-value store system based on flash storage that
scales to serve billions of key-value items on a single node. It requires
only 0.7 bytes of DRAM per entry and retrieves key/value
pairs using on average 1.01 flash reads each. SILT combines new
algorithmic and systems techniques to balance the use of memory,
storage, and computation. Our contributions include: (1) the design
of three basic key-value stores each with a different emphasis on
memory-efficiency and write-friendliness; (2) synthesis of the basic
key-value stores to build a SILT key-value store system; and (3) an
analytical model for tuning system parameters carefully to meet the
needs of different workloads. SILT requires one to two orders of
magnitude less memory to provide comparable throughput to current
high-performance key-value systems on a commodity desktop
system with flash storage.
Thus, to achieve optimal performance they construct a hybrid ensemble of key-value stores that are optimized to particular underlying media.
They start early with their performance observations, and continue this pace throughout the paper. Thus, they point out that they have a small memory requirement (0.7 bytes per key value) and only a single read from flash required:
This fits well with their observation that memory efficiency is critically important for scalable key-value stores. Their system saturated the capabilities of their hardware with 46,000 lookups per second, though this doesn’t guarantee that it will do so on newer, faster hardware (and we will get to several current KV stores.
- LogStore – uses a CPU-efficient cuckoo-hash based only upon a tag derived from the full key (20 bytes). The benefit of this approach is that it avoids accessing the flash memory if the key is not present. It is probabilistic, in that a key might not be present in Flash with some low probability.
- HashStore – uses a memory-efficient storage model for their cuckoo flash table. These stores are immutable and are stored in flash.
- SortedStore – They use an entropy-coded trie, which provides a very space efficient indexing mechanism: the average according to the authors is 0.4 bytes per key. Given that keys are 160 bits (20 bytes), which means the index is 64 bits (8 bytes). The authors provide us with an impressive list of specific optimizations they have applied to sorting data, packing indexes, intelligent (non-pointer) based addressing,
They also provide us with pretty figures to graphically lay out their data structures. First, Figure 3, which shows the in-memory cuckoo hash table:
When we move onto HashStore, they show us how the structure of the data is moved from insertion order to hash sorted order as the data is moved from memory to flash.
Of course, one of the important aspects of a key-value store is the ability to look up the data quickly. Thus, they also show us how they sort data using a prefix tree (Trie). This is a key aspect of their SortedStore:
The in-memory Trie, permits them to rapidly find the corresponding data in Flash. This is one of their key motivations for very efficient encoding of the Trie: it ensures they can store it in memory. Once they can no longer store the indices in memory the lookup system will become I/O bound (to the Flash, in this case).
Figure 6 then shows how they actually encode the length of the tree in each direction, and how it is entropy encoded:There is some serious thought that has gone into their layout. They do admit, however, that they are not providing crash tolerance: there will be data that was stored in DRAM that cannot be recovered after a crash. They do not discuss the use of NVM for achieving this. Instead, they suggest a synchronous write mechanism in which the write blocks until the log data has been written to the flash-based log.
One of the interesting aspects of their performance evaluation is the recognition that they have quite a few parameters that must be chosen. Thus, they spend more than a page of the paper just discussing the relative merits of the trade-offs involved in selecting various parameters. They consider write amplification (the need to write more than just the data being preserved), read amplification (performing reads only to discover the data that you seek is not present), memory overhead (how much DRAM is used), and update rate versus flash life time. This latter point is an interesting one to keep in mind when reading about non-volatile memory: it wears out. There is a certain amount of degradation that occurs in the underlying physical medium each time it is erased and re-used. The specific details are dependent upon the characteristics of the medium, but it is an issue to keep in mind.
The evaluation of SILT is certainly interesting. They choose not to compare to any other KV system and instead focus on the performance of their system using a variety of fairly standard loads (YCSB) as well as simple get/put key operations as micro-benchmarks. They end up claiming they they can lookup almost 45k keys per second in the slowest of their hybrid stores. When combined, they indicate they can achieve almost 58k get operations per second on a single core. As they move to multiple cores they see some further improvement (180k get operations per second for data that is already in memory).
SILT offers some interesting observations, particularly about highly efficient use of memory and novel indexing schemes.
Logic and Lattices for Distributed Programming
Neil Conway, William R. Marczak, Peter Alvaro, Joseph M. Hellerstein, and David Maier, in Symposium on Cloud Computing 2012 (SOCC ’12), October 14-17, 2012, San Jose, CA.
This is definitely a different direction than we’ve had in prior papers, though I do have an ulterior motive in presenting this particular paper – we will see it used later.
In recent years there has been interest in achieving application-level
consistency criteria without the latency and availability costs of
strongly consistent storage infrastructure. A standard technique is to
adopt a vocabulary of commutative operations; this avoids the risk
of inconsistency due to message reordering. Another approach was
recently captured by the CALM theorem, which proves that logically
monotonic programs are guaranteed to be eventually consistent. In
logic languages such as Bloom, CALM analysis can automatically
verify that programs achieve consistency without coordination.
In this paper we present BloomL, an extension to Bloom that
takes inspiration from both of these traditions. BloomL generalizes
Bloom to support lattices and extends the power of CALM analysis
to whole programs containing arbitrary lattices. We show how the
Bloom interpreter can be generalized to support ecient evaluation
of lattice-based code using well-known strategies from logic
programming. Finally, we use BloomL to develop several practical
distributed programs, including a key-value store similar to Amazon
Dynamo, and show how BloomL encourages the safe composition
of small, easy-to-analyze lattices into larger programs.
Notice they do mention key-value stores, so you have another hint on how I’ll be referring back to this work in a future post.
This tends more to the theoretical side of systems. It is not a theory paper (there just isn’t enough formalism, let alone proofs!) It has performance graphs, which you certainly expect from a systems paper, but not from a theory paper.
The driving factor behind this is the issue of distributed consistency. At a high level, “distributed consistency” is concerned with ensuring that a group of communicating computers, with some temporal separation, agree on the outcome of operations even when things go wrong. Perhaps the most famous example of distributed consistency is Paxos. These days we refer to these as consensus protocols. I generally describe there being several such: two-phase commit is certainly one of the older ones. Quorum protocols are another (e.g., weighted voting, which I described previously). Viewstamped Replication is another. These days, the popular consensus protocols are
Raft and Blockchain.
The paper starts by pointing out that monotonic consisency provides a valuable mechanism for reasoning about distributed consistency. Prior work by the authors establishes that all monotonic programs are “invariant to message reordering and retry”, a property they call confluent. This matters for distributed systems because such a system only moves forward (the operations are durable.)
They point out some weaknesses in the prior definition and motivate improving it by explaining one such obvious case that does not fit within the model (a voting quorum in a distributed protocol.)
Hence, they introduce the lattice. They do this within the context of their language (BloomL), which works on top of Ruby. I will not dwell on the details.
The authors define a bounded semijoined lattice. My reading of what they are saying is that in such a set, there is a unique element that happened first. They define this formally as a set S, with an operator (“bottom” that I don’t seem to have in my font set) that defines a partial ordering. There is a unique element ⊥ that represents the least element.
From this definition, they construct their model; the paper drops the “bounded semijoined” part of the definition and simply discusses lattices from that point forward, but it is this partial ordering property that imparts the key characteristics to their subsequent operations.
Why is this important? Because it demonstrates that these lattices – which are going to turn out to be equivalent to key operations in distributed systems – have consistency guarantees that are desirable.
The authors then turn their attention to utilizing lattices for key-value stores. They describe data structure versioning and vector clocks. Vector clocks have a property they desire for lattices: they are partially ordered. They combine this with a quorum voting protocol, to provide the distributed consensus for their system.
Figure 8 (from the paper) shows the general structure of their key-value store implementation, which is implemented in BloomL and Ruby. Their sample usage for this is a shopping cart, which they graphically describe in Figure 9 (from the paper).
As one would expect in a distributed system, the key benefit here is that there is no centralized authority deciding on the order of things. They point out that prior work argues shopping carts are non-monotonic and thus cannot be solved in a distributed systems setting. The authors point out that using the lattice structure, they achieve a monotonic ordering, which permits them to implement it without a centralized decision maker; in fact the decision maker in this case is really the client itself, as it has all the information from all the servers sufficient to complete the operation.
While a shopping cart might not be the killer application for a distributed systems technology, this paper does describe a powerful tool for providing distributed consensus in a system that can be implemented in a modest amount of code; compared to Paxos, Raft, or Viewstamped Replication, that is a significant contribution.
It does not appear to have byzantine protection, however, so if you live in a hostile environment it might not be the right protocol. Similarly, if you need stronger consistency guarantees, this might not be the best model either. But for many applications slightly relaxed consistency guarantees are often more than adequate.
We will see how this can be applied in the future.
Cache Craftiness for Fast Multicore Key-Value Storage
Yandong Mao, Eddie Kohler, Robert Morris, in Proceedings of the 7th ACM european conference on Computer Systems (Eurosys ’12), pp. 183-196, Bern, Switzerland, April 10 – 13, 2012.
In this work, the authors build a key-value system called MassTree. This is an in-memory Key-Value store (I will look at several of them, in fact). What is novel about this particular variant is the authors focus on providing a high-performance parallel access implementation and the mechanisms by which they achieve this.
MassTree support arbitrary-length keys; it assumes nothing about the keys (so they may be binary data). It uses B+-trees for storage organized into a Trie structure. The B+-tree stores a fixed slice of the key; the Trie connects the various slices together. This use is interesting, since most of the cases I’ve seen of Tries are for strings, as they do an excellent job of managing overlapping string data (prefix trees). The authors use this for binary keys. There is nothing inherently more difficult about binary versus string keys (since they are equivalent) but this choice makes the solution very flexible, as it is not particularly data dependent. This basic layout is shown in Figure 1 (from the paper).
One of the challenges with concurrent data structures is how to handle the common case – no collisions – with minimal performance overhead. The classic mutual exclusion model (or reader/writer locks) involves a certain amount of overhead, even for non-contended locks, because of the need to perform interlocked operations against shared (common) memory where the locks are maintained. The system implemented by the authors does not require any locking for readers (lookups). Updates are done with locks local to the CPU, which helps minimize the contention of typical locks.
One interesting observation is that their choice of this hybrid Trie/B+-tree structure was motivated by prior systems that struggled with performance in the presence of variable length keys. In MassTree, the rate limiting factor for queries is the cost of walking the tree. They minimize this by using “a wide fan-out tree to reduce tree-depth, prefetches nodes from DRAM to overlap fetch latencies, and carefully lays out data in cache lines to reduce the amount of data needed per node.”
Hence, my interest in this paper: these all seem to be important lessons for persistent memory as well, where the latencies are somewhat larger than for DRAM. Further, the authors are concerned about correctness. They have not looked at persistence (and recoverable consistency), so there is still further work for me to do should I investigate it further.
The authors conclude by claiming the following contributions:
First, an in-memory concurrent tree that supports keys with shared prefixes efficiently. Second, a set of techniques for laying out the data of each tree node, and accessing it, that reduces the time spent waiting for DRAM while descending the tree. Third, a demonstration that a single tree shared among multiple cores can provide higher performance than a partitioned design for some workloads. Fourth, a complete design that addresses all bottlenecks in the way of million-query-per-second performance.
Their emphasis on correctness and cache efficiency is certainly an attractive aspect of this work.
The driving considerations for their design were: efficient support of many different key distributions including binary and variable length keys, with common prefixes; fine grained concurrent access for high performance parallel access, and cache efficiency through optimal data placement and prefetching.
The authors establish the characteristics of their tree structure, including data placement. These make reasoning about their tree characteristics simpler. For example, they note that MassTree does not guarantee that it is a balanced structure. However, due to the way the tree itself is structured, they have the same algorithmic cost: O(l log n) comparisions, where l is the length of the key and n is the depth of the tree.
As a pragmatic check on their implementation, they also use a partial-key B-tree (pkB-Tree) for comparison. Despite the fact the pkB-Tree is balanced, the authors note that MassTree performs favorably well on several benchmarks. The authors go into detail about their implementation details, including the construction of border nodes and interior nodes, as well as how they lay out data (again, with an eye towards cache line efficiency).
To achieve this efficiently, they use a versioning scheme. A node has a version number. To modify the node, a given processor must update the version to indicate that it is changing the node. A reader will snapshot the version at the start of the read, and compare it at the end of the read. If it changed, the reader knows the state may have changed and can retry the read operation (essentially a variant of software transactional memory). The detailed diagram of this is shown in Figure 3 (from the paper).
The paper describes the concurrency model in the face of conflicting writers as well. By keeping their lock in the same cache line as their data, they exploit the cache coherence protocol. The authors note that lock-free operations have comparable cache behavior (e.g., compare-and-swap or link-load-store-conditional).
Indeed, much of the rest of the technical content of the paper is explaining why their approach is correct – an essential point for concurrent access systems. Without that, there really is not much point!
While brief, their discussion about value storage is interesting: their measurements are done assuming that values will be small. They state they have a scheme for managing large values as well, via a separate allocator. While they do not claim this, my observation is that a “real world” system would likely need to have some hybrid form of this.
MassTree is an in-memory key-value store. To provide persistence, they rely upon a write-behind log scheme. When used via the network interface, the writes are not guaranteed. To minimize the loss window, they choose a 200 ms timer. Thus, the log is written to disk every 200 ms. They do not evaluate this persistence model, offering it to us as an explanation that persistence is not incompatible with performance.
This last point is an interesting one, especially when considered in the context of NVM: what are the trade-offs. The authors hint at this, but do not explore this space.
For those interested, a version of the source code can be found here: https://github.com/kohler/masstree-beta
MRAMFS: A compressing file system for non-volatile RAM
Nathan K. Edel, Deepa Tuteja, Ethan L. Miller, and Scott A. Brandt in Proceedings of the 12th IEEE/ACM International Symposium on Modeling, Analysis, and Simulation of Computer and Telecommunication Systems (MASCOTS 2004), Volendam, Netherlands, October 2004.
This paper allows me to provide both a file systems paper and look at an interesting approach to byte-addressable non-volatile memory (NVM).
We have developed a prototype in-memory file system which utilizes data compression on inodes, and which has preliminary support for compression of file blocks. Our file system, mramfs, is also based on data structures tuned for storage efficiency in non-volatile memory.
One of the interesting aspects of NVM is that it has characteristics of storage (persistence) and memory (byte-addressability). Storage people are used to having vast amounts of time to do things: it is quite difficult, though not impossible, to do anything computationally with data that will be an important factor when it is combined with the overhead of I/O latency to disk drives. In-memory algorithms worry about optimal cache line usage and efficient usage of the processor, but they don’t need to worry about what happens when the power goes off.
Bringing these two things together requires re-thinking things. NVM isn’t as fast as DRAM. Storage people aren’t used to worrying about CPU cache effects on data resilience.
So mramfs looks at this from a very file systems centric perspective: how do we exploit this nifty new memory to build a new kind of RAM disk: it’s still RAM but now it’s persistent. NVRAM is slower than DIMM and hence it makes sense to compress it to increase the effective data transfer rate (though it is not clear if that really will be the case.)
I didn’t find a strong motivation for compression, though I can see the viability of it now, in a world in which we want to pack as much as we can into a 64 byte cache line. The authors point out that one of the previous systems (Conquest) settled on a 53 byte inode size. The authors studied existing systems and found they could actually compress down to 20 bytes (or less) for a single inode. They achieved this using a combination of gamma compression and compressing common file patterns (mode, uid, and gid). Another reason for this approach is they did not wish to burden their file system with a computationally expensive compression scheme.
In Figure 1 (from the paper) the authors provide a graphic description of their data structures. This depicts a fairly traditional UNIX style file system, with an inode table, name space (directories), references from directory entries to the inodes. Inodes then point to control structures that eventually map to the actual data blocks.
The actual memory is managed by the file system from a single chunk of non-volatile memory; the memory is virtually addressed and the paper points out that they don’t actually care how that mapping is achieved.
Multiple inodes are allocated together in inode blocks with each block consisting of 16 (variable length) inodes. The minimum size of a block is 256 bytes. inodes are rewritten in place whenever possible, which can lead to slack space. If an inode doesn’t fit within its existing space, the entire block is reconstructed and then written to a new block. Aftewards, the block pointer is changed to point to the new block. Then the old block is freed.
One thing that is missing from this is much reasoning about crash consistency, which surprised me.
The authors have an extensive evaluation section, comparing to ext2fs, ramfs, and jffs2 (all over RAM disk). Their test was a create/unlink micro-benchmark, thus optimizing the meta-data insertion/deletion case. They then questioned their entire testing mechanism by pointing out that the time was also comparable to what they achieved using tmpfs building the openssl package from source. Their final evaluation was done without the compression code enabled (“[U]nfortuantely, the data compression code is not yet reliable enough to complete significant runs of Postmark or of large builds…”). They said they were getting about 20-25% of the speed without compression.
Despite this finding, their conclusion was “We have shown that both metadata and file data blocks are highly compressisble with little increase in code complexity. By using tuned compression techniques, we can save more than 60% of the inode space required by previous NVRAM file systems, and with little impact on performance.”
My take-away? This was an early implementation of a file system on NVM. It demonstrates one of the risks of thinking too much in file systems terms. We’ll definitely have to do better.