Cloud

# Appendix A - Elastic Block Store (EBS)¶

EBS can be used to store HDB data, and is fully compliant with kdb+.

It supports all of the POSIX semantics required.

Three variants of the Elastic Block Service (EBS) are all qualified by kdb+: gp2 and io1 are both NAND Flash, but offer different price/performance points, and st1 is comprised of traditional drives. Unlike ephemeral SSD storage, EBS-based storage can be dynamically provisioned to any other EC2 instance via operator control. So this is a candidate for on-demand HDB storage. Assign the storage to an instance in build scripts and then spin them up. (Ref: Amazon EBS)

A disadvantage of EBS is that even if the data is read-only (immutable) a specific volume cannot be simultaneously mounted and shared between two or more EC2 instances. Furthermore, the elastic volume would have to be migrated from one instance ownership to another, either manually, or with launch scripts. EBS Snapshots can be used for regenerating an elastic volume to be copied across to other freshly created EBS volumes, which are subsequently shared around under EBS with a new instance being deployed on-demand.

Therefore, users of EBS or direct attach containing significant volumes of historical data, may need to replicate the data to avoid constraining it to just one node. You could also shard the data manually, perhaps thence accessing nodes attached via a kdb+ UI gateway.

EBS is carried over the local network within one availability zone. Between availability zones there would be IP L3 routing protocols involved in moving the data between zones, and so the latencies would be increased.

EBS may look like a disk, act like a disk, and walk like a disk, but it doesn’t behave like a disk in the traditional sense.

There are constraints on calculating the throughput gained from EBS:

• There is a max throughput to/from each physical EBS volume. This is set to 500 MB/sec for io1 and 160 MB/sec for gp2. A gp2 volume can range in size from 1 GB to 16 TB. You can use multiple volumes per instance (and we would expect to see that in place with a HDB).

• There is a further limit to the volume throughput applied, based on its size at creation time. For example, a GP2 volume provides a baseline rate of IOPs geared up from the size of the volume and calculated on the basis of 3 IOPs/per GB. For 200 GB of volume, we get 600 IOPS and @ 1 MB that exceeds the above number in (1), so the lower value would remain the cap. The burst peak IOPS figure is more meaningful for random, small reads of kdb+ data.

• For gp2 volumes there is a burst level cap, but this increases as the volume gets larger. This burst level peaks at 1 TB, and is 3000 IOPS. that would be 384 MB/sec at 128 KB records, which, again is in excess of the cap of 160 MB/sec.

• There is a maximum network bandwidth per instance. In the case of the unit under test here we used r4.4xlarge, which constrains the throughput to the instance at 3500 Mbps, or a wire speed of 430 MB/sec, capped. This would be elevated with larger instances, up to a maximum value of 25 Gbps for a large instance, such as for r4.16xlarge.

• It is important note that EBS scaled linearly across an entire estate (e.g. parallel peach queries). There should be no constraints if you are accessing your data, splayed across different physical across distinct instances. e.g. 10 nodes of r4.4xlarge is capable of reading 4300 MB/sec.

Kdb+ achieves or meets all of these advertised figures. So the EBS network bandwidth algorithms become the dominating factor in any final calculations for your environment.

For consistency in all of these evaluations, we tested with a common baseline using an r4.4xlarge instance with four 200-GB volumes, each with one xfs file system per volume, therefore using four mount points (four partitions). To show the scale to higher throughputs we used an r4.16xlarge instance with more volumes: eight 500-GB targets, (host max bandwidth there of 20 Gbps, compared with max EBS bandwidth of 1280 MB/sec) and we ran the comparison on gp2 and io1 versions of EBS storage. For the testing of st1 storage, we used four 6-TB volumes, as each of these could burst between 240-500 MB/sec. We then compared the delta between two instance sizes.

## EBS-GP2¶

function latency (mSec) function latency (mSec)
hclose hopen 0.004 ();,;2 3 0.006
hcount 0.002 read1 0.018

## EBS-IO1¶

function latency (mSec) function latency (mSec)
hclose hopen 0.003 ();,;2 3 0.006
hcount 0.002 read1 0.017

## EBS-ST1¶

function latency (mSec) function latency (mSec)
hclose hopen 0.003 ();,;2 3 0.04
hcount 0.002 read1 0.02

## Summary¶

Kdb+ matches the expected throughput of EBS for all classifications, with no major deviations across all classes of read patterns required. EBS-IO1 achieves slightly higher throughput metrics over GP2, but achieves this at a guaranteed IOPS rate. Its operational latency is very slightly lower for meta data and random reads. When considering EBS for kdb+, take the following into consideration:

• Due to private-only presentations of EBS volumes, you may wish to consider EBS for solutions that shard/segment their HDB data between physical nodes in a cluster/gateway architecture. Or you may choose to use EBS for locally cached historical data, with other file-systems backing EBS with full or partial copies of the entire HDB.

• Fixed bandwidth per node: in our testing cases, the instance throughput limit of circa 430 MB/sec for r4.4xlarge is easily achieved with these tests. Contrast that with the increased throughput gained with the larger r4.16xlarge instance. Use this precept in your calculations.

• There is a fixed throughput per GP2 volume, maxing at 160 MB/sec. But multiple volumes will increment that value up until the peak achievable in the instance definition. Kdb+ achieves that instance peak throughput.

• Server-side kdb+ in-line compression works very well for streaming and random 1-MB read throughputs, whereby the CPU essentially keeps up with the lower level of compressed data ingest from EBS, and for random reads with many processes, due to read-ahead and decompression running in-parallel being able to magnify the input bandwidth, pretty much in line with the compression rate.

• st1 works well at streaming reads, but will suffer from high latencies for any form of random searching. Due to the lower capacity cost of st1, you may wish to consider this for data that is considered for streaming reads only, e.g. older data.