FAQs about the performance of the internal FC-AL of the disk array

Winter melon head Q & A: disk array internal FC-AL performance problems winter melon head Source: IT expert network

Q:

Most disk arrays are connected to all disks through a 1/2/4 FC-AL ring consisting of two controller backend ports.

The FC-AL arbitration ring Agreement stipulates that at the same time only two devices can transmit data, that is, in a FC-AL ring, a back-end port of the Controller acts as the initiator, A hard disk on the ring acts as the destination. At one time, the backend port sends a data read/write command. Only one hard disk can respond to the command and transmit data.

Does that mean that the overall performance of a FC-AL ring depends on the read/write performance of a hard disk? For a hard drive with a 15 K speed, the sustained read/write bandwidth is less than 70 MB/S, And the iops is less than 400. Like IBM ds4800, EMC CX-80 and so on, a total of 4 rings, the backend performance is only 280 Mb/s, 1600 iops? I don't know if there is any other way, except to speed up with cache in the array, or is there a non-industrial standard FC-AL inside the disk array?

A:

This problem is very good and classic. To explain this problem, you need to understand three points:

1. The fcal transmission channel does allow data transmission only through two exclusive channels at the same time.

2. If the Controller has sufficient IO requests, it will never idle the channel and make full use of the bandwidth.

3. external transfer rate and internal transfer rate of the disk. When multiple devices exist on the fcal loop, the Controller's polling policy makes full use of the bandwidth, the entire system is represented as a big virtual device that is always reading and writing data, rather than tracing (also described on page 53rd of "big talk storage"). Once a device needs to track, so that other devices can transmit data to make up for the wasted time slot, so the overall system can exert the internal transmission rate of a single device.

The following is a detailed summary:

When multiple devices exist in the fcal loop, after the Controller initiates Io to the device, it takes some time to find the device. During this period, the Al ring is released, at this time, the controller can still initiate Io to another device, which is similar to sending all the commands that should be done first. When a device completes the tracing request and returns the data to the Controller, it is often because multiple devices are in the accumulation status, that is, they have finished their work and are ready to make a delivery gap. At this time, they can only queue one by one, and everyone is fighting. After understanding this, let's look at it.

　　Contradictions about iops values:

There is a conflict between iops and throughput. In the environment where iops is concerned, Io size is usually relatively small, because only a small size will not be filled with bandwidth to reach the bottleneck, so io size needs to be relatively small to reach a high iops. In this case, after the controller sends an I/O Request to the device, multiple devices will receive data transmission opportunities one by one based on the Accumulation status. Due to the small Io size, therefore, each time the data is transmitted, it will soon end. In this way, an IO will be completed quickly, and the next device will soon complete after the I/O of the previous device is completed, because it is out of accumulation, the data to be returned has already been prepared to be sent in the cache.

In this way, the overall system is represented as a virtual device that is always performing I/O and does not need to be searched. If there is only one device except the Controller on the Al ring, the ring must wait for it to seek the channel, because no other device can work in the Al ring within the time slot.

However, the effect of making up for the time slot of the Al ring is not that the more devices, the better. Different designs and products have their own optimal number of devices. The current experience is 64, that is, half of the total loop capacity. If the total capacity exceeds this value, the performance will not be improved or even decreased.

We can draw another conclusion, that is, a slow device, for example, a device with a long track time, the more slow the device is, the larger the overall increase it brings after forming an Al loop, the faster the device is, A high-specification device, after forming an Al loop, has limited performance improvement. This is the effect that the Al ring or other shared bus/ring methods make up for the time slot produced by the device's own processing.

　Conflicts about throughput/bandwidth values:

After the above description, we have a good understanding of the underlying mechanism of the shared bus/ring transmission mode and its utility. I/O size is often very large in an environment that focuses on and pursues high throughput, that is, making full use of bandwidth, so that the channel bandwidth can be reached at a lower iops value.

In this case, the upper layer sends large consecutive Io segments, so that the tracing time of devices on the Al ring can be greatly reduced, which makes the devices faster out of the accumulating state, prepare to send data externally. We know that the external transfer rate of the disk is interrupted due to the constant internal change, which causes the value to be approximately 20 times lower than the internal transfer rate. The utility of the Al ring is to make up for the time slot wasted by seeking for channels. Therefore, the external throughput of the overall system is increased, thus resolving this contradiction.