Our study reveals the induction of channels allowing for the nuclear egress of the progeny viruses across the host chromatin. Moreover, this work used a new combination of methods in the study of virus-cell interactions.Language production appears to be a largely incremental process: speakers plan an utterance as they are producing it, simultaneously integrating multiple sources of information . One apparent effect of this incrementality is availability effects in language production: the fact that speakers will often choose to produce words which are easily accessible or available to them earlier in an utterance, or to include such words when they are optional, or even to use a highly-available word in place of a more communicatively accurate but less available word . The fact that available words tend to go earlier has been attributed to a greedy, ‘easy-first’ language production strategy . Here I present an account of availability effects within a computational-level model of language production based on a recently developing theory of the complexity of action selection from the fields of computational neuroscience and information theory. This theory, the rate–distortion theory of control , holds that actions are selected to maximize value subject to constraints on the use of information. The theory originates in the economics literature where it operationalizes bounded rationality , and it has been developed and applied in the literature on physics, robotics, optimal control,25 liter pot computational neuroscience, reinforcement learning, cognitive psychology, and linguistics .
RDC uses the mathematical theory of lossy compression to impose informational constraints on the perception–action loop. It has also been termed rational inattention and policy compression. I develop a proof-of-concept model of language production within the RDC framework, based on an informational constraint identifiable as a channel capacity limit on cognitive control . I show that this model provides an account of availability effects in language production, and I validate this account by examining experimental data from two previous sets of experiments: Levy & Jaeger on relative clause complementizers in English, and Zhan & Levy on noun classifier choice in Mandarin Chinese. In contrast with existing models of language production which are primarily situated at Marr’s algorithmic level of analysis or at more concrete levels, the RDC model is at the computational level: it directly describes the inputs, outputs, and goals of the language production system, without committing to an algorithmic implementation. The high level of abstraction means that it is possible to see how simple underlying computational constraints give rise to a variety of different behaviors. Hard disks consume a significant amount of power. In general purpose computing, hard disks can be responsible for as much as 30% of a system’s power consumption. This percentage will only increase as current CPU trends lean toward increasing the number of cores versus the single core clock rate, hard disks use faster rotational speeds, and multiple hard disks per system become more prevalent. In large storage systems, hard disks can dominate system power consumption: 86% and 71% of the total power consumption in EMC and Dell storage servers, respectively.
As a result, there are several motivations to decrease the power consumed by hard disks, from increasing battery lifetime in mobile systems to reducing financial costs associated with powering and cooling large storage systems. To reduce hard disk power consumption, spin-down algorithms are used, which put a disk in a low-power mode while it is idle. In a low-power mode, such as standby, the platter is not spinning and the heads are parked, reducing power consumption. Researchers have proposed several spin-down algorithms, which are very efficient at reducing hard disk power consumption. These algorithms are typically time-out driven, spinning down the disk if the time-out expires before a request occurs. Adaptive spin down algorithms vary the time-out value relative to request inter-arrival times. They are very effective and approach the performance of an optimal off-line algorithm which knows the inter-arrival time of disk requests a-priori. Although spin-down algorithms are effective at reducing hard disk power consumption, pathological workloads can completely negate a spin-down algorithm’s power saving benefit, prematurely causing a disk to exceed its duty cycle rating, and significantly increasing aggregate spin up latency. Such pathological workloads, which periodically write to disk, are not uncommon. Both Windows and UNIX systems exhibit such behavior. For example, Figure 1 shows the periodic disk request pattern of an idle Windows XP system. In UNIX systems, applications such as task schedulers , mail clients, and CUPS periodically write to disk. Upcoming hybrid disks will place a small amount of flash memory logically next to the rotating media, as shown in Figure 2. The first hybrid disks will either have 128MB or 256MB of NVCache in a 2.5 in form factor 1.
A host can exploit the NVCache to achieve faster random access and boot time because it has constant access time throughout its block address space as shown in Figure 3, while rotating media suffers from rotational and seek latency. Access time for this particular device is roughly equal to c+bs÷bs off, where c is a 2.2ms constant overhead, bs is the desired blocksize, and bs off is 4KB. In addition to the potential performance increase, hybrid disks can potentially yield longer spin-down durations— the NVCache can service I/O while the disk platter andarm are at rest, such as the write requests from Figure 1. Note that because flash memory is non-volatile, NVCachestored data is persistent across power loss. To exploit the underlying media characteristics of hybrid hard disks for improved power management, we present four enhancements to increase power savings, reliability, and reduce observed spin-up latency: Artificial Idle Periods that extend idle periods relative to observed I/O type; a Read-Miss Cache that stores NVCache read-miss content in the NVCache itself; Anticipatory Spin-Up that spins the rotating media up in anticipation of an I/O operation not serviceable by the NVCache; and, NVCache WriteThrottling that limits the reliability impact imposed on the NVCache because of I/O redirection.We now present an overview of a hybrid disk and how its NVCache can be managed by a host operating system using a modified set of ATA commands,25 liter plant pot according to the T13 specification for hybrid disks. The four enhancements are presented in Section 3. Sectors stored in the NVCache are either pinned or unpinned, which when referred to as a collection are known as the pinned and unpinned set, respectively. The host manages the pinned set, while the disk manages the unpinned set. Hybrid disks will also have a new power mode, NV Cache Power Mode, which can be set and unset by the host. In this mode, I/O is directed to the NVCache unpinned set while the disk “aggressively” tries to keep the rotating media spun-down. Defining and implementing “aggressive” is left to the drive vendor’s discretion. Although the hybrid disk controls the spin-down policy, the host controls the minimum time rotating media must remain spinning after a spin-up, providing the host with some control over the underlying spin-down algorithm. The host controls I/O to the NVCache pinned set. Sectors can be pinned in the NVCache, and pinned sectors can be removed or queried.
The pinned attribute feature is intended to increase random access performance, although it can also facilitate better power management. The host can flush a specific amount of unpinned content to rotating media to make room for more pinned sectors. However, pinned sectors cannot be evicted to create unpinned space. Addressing multiple sectors at a time is possible using an extents-based mechanism called LBARange Entries. A host can also specify the source when adding pinned sectors to the NVCache: host or rotating media, by setting a Populate Immediate bit. This capability gives the host control over NVCache functionality: better random access or spin-down performance. A host has additional control over a hybrid disk. It can query the disk for spin-up time, read/write NVCache throughput, and the maximum pinnable sectors.We now discuss the mechanism in which a host can leverage a hybrid disk to provide power management functionality. The host controls rotating media state with traditional power management commands, and NVCache commands to manage the pinned set. Fine-grain spin-down algorithms can be implemented because the host is informed when rotating media power state changes occur. While the rotating media is spun-down, the host should use the NVCache store, query, and read commands to redirect I/O to the NVCache. If the NVCache does not have the requested read data, or it is full before a write, the rotating media must be spun-up, the request satisfied, and NVCache content flushed to disk using both pinned set removal and traditional disk I/O commands. The host can put the disk in standby mode again when the spin-down algorithm deems it desirable to do so. We assume this method because it provides us with complete control over a hybrid disk, allowing us to implement a fine-grain adaptive spin-down algorithm and I/O subsystem enhancements to exploit a hybrid disk’s media characteristics. Alternatively, a host could rely on the NV Cache Power Mode to provide all aspects of power management. However, there are several limitations with this approach: the minimum high-power time is not dynamic, it assumes the disk controller implements the correct spin-down policy, and the NVCache may not be a suitable I/O destination for certain workloads. A host could implement its own coarse-grain spin-down algorithm, by repeatedly entering and exiting the NV Cache Power Mode, recording I/O response times to implicitly infer when the rotating media is spun-up. In this way, a host can utilize its own spin-down algorithm, but still has no control over NVCache management. Note that we omit pinned and unpinned references for the remainder of this work as we no longer refer to the unpinned set.Mean Time To Failures and Mean Time Between Failures are widely used metrics to express disk reliability. However, disk manufacturers also provide a duty cycle rating. Duty cycle rating is the number of times rotating media can be spun down before the chances of failure increase to more than 50% on drive spin-up. When controlling a disk’s power state with a spin-down algorithm, the duty cycle metric is potentially more important than either MTBF or MTTF because a spin-down algorithm results in an accelerated consumption of duty cycles. In addition to duty cycles, hybrid disks also have flash memory reliability—flash memory blocks have a rated number of erase cycles they can endure before errors are expected. Today’s hard disks generally consume 5–10 times more energy while in active than in standby mode. As a result, adaptive spin-down algorithms are very aggressive— it is more efficient to spin-down after only a few seconds of idle time. With such short idle times the number of duty cycles increases dramatically. Duty cycle terminology varies, depending on drive class and technology. Typically, 3.5 in drives refer to duty cycles as Contact Start/Stop Cycles , where the head comes to rest on a landing zone on the platter during a power-down. An alternate technology, ramp load/unload, is typically used in notebook drives, where the head comes to rest off the side of the platter. Drives using CSS technology have duty cycle ratings in the range of 50,000, while drives with ramp load/unload technology are in the range of 500,000, mostly due to reduced stiction effects. With current compact flash specifications, the number of erase operations per block is typically rated at 100,000 with 256KB sized erase blocks. A hybrid disk containing a 256MB NVCache can keep its rotating media spun-down while up to 256MB of data is written to it. With optimal wear-leveling and a write-before-erase architecture, a 256MB device can endure over 100 million erase operations before becoming unerasable. An optimal wear-leveling algorithm spreads all writes across the entire device’s physical address space while write-before erase architecture always writes data corresponding to the same LBA to an empty physical location to ensure data corruption does not occur on a bad overwrite. By exceeding the block erase rating, flash memory blocks may become unerasable, but are still readable. To a host, a hybrid disk with unerasable NVCache blocks should appear as a traditional disk.Spin-down algorithms which control the power state of traditional hard disks are efficient at reducing disk power consumption. There is little room for improvement of such algorithms, which dynamically adjust to the most power efficient time-out using machine learning techniques.