Llama - Low Latency Adaptive Media Algorithm Tomasz Lyko, Matthew Broadbent, Nicholas Race Lancaster University, United Kingdom {t.lyko, m.broadbent, n.race}@lancaster.ac.uk Mike Nilsson, Paul Farrow, Steve Appleby British Telecommunications plc, United Kingdom {mike.nilsson, paul.farrow, steve.appleby}@bt.com Abstract—In the recent years, HTTP Adaptive Bit Rate (ABR) streaming including Dynamic Adaptive Streaming over HTTP (DASH) has become the most popular technology for video streaming over the Internet. The client device requests segments of content using HTTP, with an ABR algorithm selecting the quality at which to request each segment to trade-off video quality with the avoidance of stalling. This introduces high latency compared to traditional broadcast methods, mostly in the client buffer which needs to hold enough data to absorb any changes in network conditions. Clients employ an ABR algorithm which monitors network conditions and adjusts the quality at which segments are requested to maximise the user’s Quality of Experience. The size of the client buffer depends on the ABR algorithm’s capability to respond to changes in network conditions in a timely manner, hence, low latency live streaming requires an ABR algorithm that can perform well with a small client buffer. In this paper, we present Llama - a new ABR algorithm specifically designed to operate in such scenarios. Our new ABR algorithm employs the novel idea of using two independent throughput measurements made over different timescales. We have evaluated Llama by comparing it against four popular ABR algorithms in terms of multiple QoE metrics, across multiple client settings, and in various network scenarios based on CDN logs of a commercial live TV service. Llama outperforms other ABR algorithms, improving the P.1203 Mean Opinion Score (MOS) as well as reducing rebuffering by 33% when using DASH, and 68% with CMAF in the lowest latency scenario. I. I NTRODUCTION Video streaming has been the largest service on the internet for years and it continues to grow, this includes live video streaming, as seen in the most recent Cisco report [1]. At the moment, many video streaming services use HTTP Adaptive Streaming technologies such as MPEG Dynamic Adaptive Streaming over HTTP (DASH) [2], where content is split into short segments (usually from 2-10 seconds), encoded at multiple bit rates and then hosted on a standard HTTP server. A manifest file is created that indicates the encoded bit rates and where the content can be obtained. The client requests the manifest file, then makes HTTP requests for consecutive segments of content at bit rates selected by an Adaptive Bit Rate (ABR) algorithm, which takes into account measurements of the network conditions to maximise the Quality of Experience (QoE) of the viewer. When using DASH for live content services, the client can only request segments after they become available on the server, as indicated by the manifest file. The use of DASH for live video streaming usually causes high end to end latency compared to traditional broadcast methods due to client buffering to queue segments prior to playback to allow time for the ABR algorithm to adapt changes in network conditions. If the client buffer is depleted, the playback will stall, and the QoE will be reduced. For live streaming, the maximum amount of data that could be queued in the client buffer depends on the end to end latency, as only data that has been captured and encoded but not yet presented to the user can be buffered. For example, if the client is playing content three segment durations behind live, the maximum client buffer fill is three segments. Recently, the Common Media Application Format (CMAF) [3] has been standardised, enabling segments to be divided into chunks to aid low latency live streaming. In this paper, we present Llama, an ABR algorithm specif- ically designed to operate at low buffer levels to offer better QoE for low latency live streaming scenarios. It can be used with HLS, DASH and CMAF. We compare the performance of Llama against four popular ABR algorithms in terms of multiple QoE metrics, across multiple client settings, and in various network scenarios based on CDN logs from a com- mercial live TV service. In order to perform such an extensive evaluation, we have developed and verified a simulation model which supports live DASH and CMAF. II. BACKGROUND AND RELATED WORK A. ABR Algorithms Bentaleb al. [4] published a survey of ABR algorithms used in HTTP Adaptive Streaming. They outlined the main goal of an ABR algorithm is to maximise viewer QoE. This involves trying to maximize the average video quality, while trying to minimize the number of rebuffering events, the time spent in the rebuffering state, and the frequency of changes of video quality. They note that most of these goals are in competition with each other, and therefore require a reasonable trade-off. This survey classifies client-side ABR algorithms into the following five classes: available bandwidth-based, playback buffer-based, proprietary solutions, mixed, and Markov Deci- sion Process (MDP)-based. Available bandwidth-based ABRs select video quality based on estimates of the available band- width, usually by calculating the throughput of the previously fetched segment(s). Playback buffer-based ABRs select video quality based on the buffer level alone, usually only selecting a video quality once its buffer level threshold has been reached. Proprietary solutions include ABR algorithms from adaptive streaming solutions and player implementations, such as Apple HTTP Live Streaming (HLS) [5]. Mixed ABR algorithms