A performance Brokerage for Heterogeneous Clouds

4.4 Compute as a Commodity Market

If we consider again how providers define instance types, by specifying some number of vCPUs, a quantity of RAM, an amount of local storage and a per vCPU compute rating, then it becomes clear why the notions of commodity naturally arise when discussing Clouds. Each instance type may be considered akin to a commodity class as instances of the same type are interchangeable. Further, whilst instances of the same type may run on hosts with different CPUs, so long as they meet the specified criteria, as defined by an ECU or GCEU for example, they are considered equivalent. This does of course leave room for a degree of variation between equivalent instances, particularly so if the criteria is defined in terms of a minimum.

Extending this commodity view of computing, it is a small step to imagine different providers selling the same types of instances which meet a common agreed specification and, further, doing so through an organized marketplace. As all instances will meet a common specification, providers will compete on the basis of price, and users will consider them to be fungible, meaning an instance from one provider can be swapped with an instance from another. Buyya et al. (2009), Garg et al. (2013) Cartlidge (2014), and Weinman (2015) have suggested Cloud marketplaces, organized along the lines of extant commodity exchanges, through which Cloud resources can be traded.
Exchanges bring a number of advantages, such as price transparency, continuous price discovery and liquidity (high volumes of activity). The latter is ensured through the presence of market makers, market participants who continuously post bids and asks i.e. buy at and sell at prices. Liquidity means prices are continuously discovered, and whilst prices will trend up and down, price deviations (changes from one price to the next) are typically small, helping to reduce risk for both buyers and sellers. A compute exchange would have considerable benefits to users, with liquidity alleviating problems such as availability and provider failure. In addition, dynamic pricing means that providers can charge more when demand is high, whilst users can take advantage when prices are low.
Exchanges typically allow for commodities to be purchased ‘on the spot’ for immediate delivery, or via a futures contract. In the latter, a price is agreed when the contract is purchased but delivery dates are in the future. Future contracts on exchanges are also standardised, allowing them to be traded, and include a quantity of good to be delivered as well as a delivery date. Commodity options, which give the owner the right but not the obligation to buy/sell the underlying commodity, are typically traded off-exchange, in the so-called Over the Counter (OTC) market.
OTC markets allow for the trading of both bespoke (non-standard) or standardised assets in non-standard quantities. These markets are unregulated and do not specify how trading takes place or require the publication of pricing information, are popular due to their ability to offer customised products, and are commonly used by traders for the hedging of specific risks. Unlike exchanges, which act as a counter party to all trades, OTC markets are bi-lateral and so have greater credit risk, and without visible pricing across the market there is a lack of price transparency. Further, without market makers, OTC markets lack liquidity. A report by the IMF (Dodd and Mills, 2008) states that the market for collaterised debt obligations (CDOs) quickly became illiquid during the financial crisis of 2007-2008, leading to problems at major investment banks such as Lehman Brothers.
Notably, whilst Buyya at al. (2009) envisaged the development of a Cloud exchange whereby participation is through bids and asks, a key component of their proposed market is SLAs: CSBs and providers engage in an SLA negotiation process which may or may not succeed. This type of negotiation is not found on exchanges, but is a feature of OTC markets. Similarly, Garg et al. (2013) presuppose a Cloud market whereby providers will respond to a request for quotes (RFQ) from brokers, which is also a typical feature of OTC markets.
Weinerman (2015) proposes that compute exchanges, in addition to a spot market, may offer compute derivatives which allow users to hedge against future price movements as they seek to manage workload execution costs. Future contracts are also beneficial to providers as they guarantee some level of future income, thus making investment decisions easier. However, managing workload execution costs will require performance to be specified as part of the contract. Notably, the Chicago Mercantile Exchange (CME), responsible for the largest derivatives exchange in the world, explored derivative contracts (futures and options) based on compute resources (Cohen, 2013), although to date no such contracts are traded on the exchange.
On the majority of the world’s financial exchanges trading typically takes place through a so-called continuous double auction (CDA) whereby bids (a buy at price) and asks (a sell at price) are continuously posted onto a central limit order book (CLOB), and all market participants can see both sides of the book. A CLOB has 2 sides, a bid side and an ask side, with the ask side in ascending price order and the bid side in descending. Orders with the same price are ordered by time, with existing orders taking precedence, and the arrival of a new order triggers a re-ordering. By the best ask/bid we mean the ask/bid that is at the top of the order book. The bid-ask spread is the difference between the best bid and the best ask, that is, the difference between the highest bid price and the lowest ask price.
By order book execution we mean the process of matching bids and asks, which we describe now. If the ask price is below the best bid price the orders are matched. If the bid quantity is sufficient to fulfil the ask, then the ask is removed and the bid is updated, with the bid quantity being reduced by the amount sold. However, if the bid quantity is less than the ask quantity the bid is removed and the ask is updated. The ask is then matched against the new best bid. This process continues until either the ask is fulfilled, and so removed, or there are no bids with which to be matched due to price increases on the bid side. In the latter case the order has been partially fulfilled and an ask order will remain on the order book for some period of time before being removed. The current market price, or spot price, is the price at which the last transaction occurred.
Exchanges typically support a variety of different order types, so as to meet a variety of requirements, with different types being distinguished by (1) the length of time the order is good for (2) whether or not partial fulfilment is acceptable and (3) whether or not the price is limited or if the current spot price is acceptable. For the last point, a limit order specifies a bid/ask price and so limits the price that will be accepted. For example a limit bid price of £1 will not be matched with any ask offering above this. A market order however will accept the current spot price, but must be executed immediately, meaning it is executed as far as possible and then removed.
Marketplaces are typically modeled as a continuous double auction (CDA) and it is commonplace to make simplifying assumptions regarding the types of orders and the quantities being sold. For example, Cliff and Bruten (1997) and Gode and Sunder (1993) limit the quantity of goods being sold to one unit and so there is no need for multiple counter parties to fulfill a trade. Further, bids/asks are executed immediately. Gode and Sunder randomly select a buyer and a seller (from those with sufficient quantity to buy/sell) and generate a bid and an ask. If the prices cross then a transaction occurs; if not then both orders are removed. In this way the order book is randomly generated at each point in time.
Similarly, Cliff and Bruten (1997) chooses a buyer/seller in each time slice to quote a bid/ask, and all sellers/buyers who are active and wish to do so will accept the quote – if a quote is accepted by multiple participants then one of them is selected at random. Gode and Sunder (1993) have traders randomly generate prices whenever chosen - so called Zero Intelligence Constrained¹⁹ (ZI-C) traders – whilst Cliff implements Zero Intelligence Plus (ZIP) which allows traders to implement a pricing strategy.
In a series of Nobel Prize winning behavioral experiments, Smith (1962) shows that in a CDA market consisting of only 12 sellers and 11 buyers, whereby participants are provided with trading guidelines, the market price rapidly approaches the theoretical equilibrium price. This is notable as it was the first work to demonstrate how an equilibrium price may be obtained. Equilibrium pricing is important as it is the price at which supply and demand are equal, and so there is neither excess supply nor demand, which Economists refer to as an efficient allocation. Less formally, Cliff describes lack of equilibrium pricing as “…somebody somewhere is being ripped off ”.
Why CDAs produce equilibration is unknown. As noted, Gode and Sunder (1993) consider so-called Zero Intelligence Constrained (ZI-C) traders, these are traders which are given a range of prices (a different range per trader) and submit randomly chosen bids/asks within their set ranges, and if they are unsuccessful in a bid/ask they simply submit a randomly chosen new offer. Due to simplicity of implementation and ability to find equilibrium prices under certain conditions they are widely used as trading agents. Indeed, it was claimed that ZI-C traders participating in a CDA produce known theoretical equilibrium price, and so it is solely the properties of the CDA that produces the equilibrium. However, Cliff and Bruten (1997) shows that ZI-C traders only produce equilibrium prices under certain conditions. In an electronic replica of the Smith experiments he demonstrated that a market of ZIP traders operating under a CDA quickly reaches equilibration.
Interestingly, Cartlidge (2014) considers the benefits of changing the structure of the current ARIM market, where only sellers can post prices, to a CDA where both buyers and sellers can post. The former is referred to as a retail market. He introduces ZIP buyers, as traders who can post asks (sale prices) but cannot post bids. He shows that the retail market under-prices instances, as compared to known theoretical equilibrium prices by up to 10%, confirming earlier results by Smith (1962) on retail markets. However, changing the ARMA to a CDA, by allowing ZIP traders to post bids and asks, means the market rapidly equilibrates.
In the next section we discuss pricing approaches that are specific to the Cloud, so as to inform how our performance broker may price instances.