Modeling and optimizing IO throughput of multiple disks on a

Abstract For a wide variety of computational tasks, disk I/O continues to be a serious obstacle to high performance. The focus of the present paper is on systems that use multiple disks per SCSI bus. We measured the performance of concurrent random I/Os, a

Modeling and optimizing I/O throughput of multiple disks on a bus (summary)Rakesh Barve Elizabeth Shriver

y

Phillip B. Gibbons Je

y

Bruce K. Hillyer

y

Yossi Matias

z

rey Scott Vitter

Abstract For a wide variety of computational tasks, disk I/O continues to be a serious obstacle to high performance. The focus of the present paper is on systems that use multiple disks per SCSI bus. We measured the performance of concurrent random I/Os, and observed bus-related phenomena that impair performance. We describe these phenomena, and present a new I/O performance model that accurately predicts the average bandwidth achieved by a heavy workload of random reads from disks on a SCSI bus. This model, although relatively simple, predicts performance on several platforms to within 12% for I/O sizes in the range 16{128 KB. We describe a technique to improve the I/O bandwidth by 10{20% for random-access workloads that have large I/Os and high concurrency. This technique increases the percentage of disk head positioning time that is overlapped with data transfers, and increases the percentage of transfers that occur at bus bandwidth, rather than at disk-head bandwidth. Introduction We study the performance of workloads consisting of read requests directed to disks that share a SCSI bus. These workloads are relevant to certain multimedia, database, and scienti c computing applications that use external memory and out-of-core algorithmic techniques (e.g., 6, 1]). For our experiments, the requests are generated by one process per disk. Each process iterates the following steps: (1) generate a random block address, (2) record a timestamp, (3) issue a seek and a read to the raw disk, (4) record a timestamp when the read request completes, (5) return immediately to step 1. We measured requests ranging from 16 KB to 128 KB, on four hardware con gurations. (1) 1{4 Seagate Cheetah disks, Sun Ultra-1, Solaris 2.5.1. (2) 1{4 Seagate Cheetah disks, Sun Sparc-20, Solaris 2.5. (3) 1{4 Seagate Barracuda disks, DEC AlphaStation 600 5/266, Digital UNIX 4.0. (4) 1{7 Seagate Wren-7 disks, Sun Sparc 20, Solaris 2.5. A number of performance models exist for disks. The analytic disk model of 4] captures bus e ects only in the singledisk case. The Pantheon disk simulator 3] incorporates bus contention and other bus e ects, but no results have been published that describe the idle periods and head-limitedDuke University, visiting Bell Labs. Sciences Research Center at Bell Labs, contact shriver@http://. z Tel-Aviv University. Work performed at Bell Labs.y Information

bus transfers that we observe. The method for approximating the throughput of multiple disks on a SCSI bus in 2] sums the seek time, rotational latency, and transfer time, and derates the performance by a contention factor derived from a general queuing model.

Rounds In our experiments, we typically observe that all disks receive a request, then all disks transmit data back to the host before any disk rec

eives another request. We use the term rounds for this periodic convoy behavior. We observed rounds for workloads of random accesses as well as when the accesses on each disk have spatial locality. We were surprised to see rounds. Since the host has the highest SCSI priority, one would expect that soon after a disk completes one request, the host would seize the bus to send another request to that disk, thereby keeping the bus and all the disks busy. Rounds could arise if the operating system kernel implements a fairness policy that forcibly balances the number of requests sent to each disk during periods of heavy I/O, rather than sending requests to disks as soon as possible. The current literature does not discuss rounds as we observed them. In fact, 5] states that even in cases when the load is symmetrically distributed and balanced, one process can monopolize the disks while others starve. Previous models (e.g., 2]) gave prediction errors greater than 100% for our workloads. We develop a model below that gains accuracy from three factors: It accounts for idle times and overlaps caused by rounds, it distinguishes between the transfer rate from a disk platter through the disk head and the (much higher) rate from a disk's cache to the host, and it considers the e ect of a relatively obscure disk control parameter called the fence or bu er full ratio. The fence determines the time at which the disk will begin to contend for the SCSI bus. A minimum fence (i.e., 0) causes the disk to contend after reading the rst sector of data into the disk's internal cache. A maximum fence (i.e., 255) causes the disk to wait until almost all of the requested data has accumulated in the disk cache. When the bus is idle, a low fence value starts sending data sooner, but more of the transfer occurs at the head-limited bandwidth. Model We developed a model that predicts the read response time for a single disk and for multiple disks on a bus. Because of space limitations, we present only the more interesting multiple disk equations here. In each round, one request for B bytes is served from each of D disks. When the fence is 0, the idle time at the beginning of the round is the expected minimum positioning time over all disks, denoted MPT(D), plus the overhead time for a SCSI command, denoted Os. The rst disk to respond transmits one sector at bandwidth BW, and continues at BW for an amount of data denoted L(B ). Then the disk disconnects,bus rot

Abstract For a wide variety of computational tasks, disk I/O continues to be a serious obstacle to high performance. The focus of the present paper is on systems that use multiple disks per SCSI b …… 此处隐藏：10072字，全部文档内容请下载后查看。喜欢就下载吧 ……